Production increases in Asia, however, have been staggering: meat production has increased fold since Absolute increases in production in other regions have also been substantial, with output in all regions with exception to the Caribbean which approximately tripled growing more than 5-fold over this period. However, the distribution of meat types varies significantly across the world; in some countries, other meat types such as wild game, horse, and duck can account for a significant share of total production.
Although production of all major meat types have been increasing in absolute terms, in relative terms the share of global meat types have changed significantly over the last 50 years. In , poultry meat accounted for only 12 percent of global meat production; by its share has approximately tripled to around 35 percent.
In comparison, beef and buffalo meat as a share of total meat production has nearly halved, now accounting for around 22 percent. In the chart we see the global production of cattle beef and buffalo meat. Globally, cattle meat production has more than doubled since — increasing from 28 million tonnes per year to 68 million tonnes in Global production of poultry meat has increased rapidly over the last 50 years, growing more than fold between Global trends in poultry production are shown in the chart.
China and Brazil are also large poultry producers at 18 and 13 million tonnes, respectively. Collectively, Europe is also a major poultry producer with an ouput in of approximately 19 million tonnes — just below output of the United States.
China dominates global output, producing just short of half of total pigmeat in Increases in Chinese pigmeat production have been rapid, growing around fold from 1. Global population has undergone rapid growth , especially in the second half of the 20th century; we may therefore also expect the rapid growth in total meat production as explored in the sections above.
But how has meat consumption changed on a per capita basis? In the chart we see a global map of per capita meat excluding seafood and fish consumption, measured in kilograms per person per year. As a global average, per capita meat consumption has increased approximately 20 kilograms since ; the average person consumed around 43 kilograms of meat in This increase in per capita meat trends means total meat production has been growing at a much faster than the rate of population growth.
The direction and rate of change across countries has highly variable. Growth in per capita meat consumption has been most marked in countries who have underwent a strong economic transition — per capita consumption in China has grown approximately fold since ; rates in Brazil have nearly quadrupled. The major exception to this pattern has been India: dominant lactovegetarian preferences mean per capita meat consumption in was almost exactly the same as in at less than 4 kilograms per person.
Meat consumption is highest across high-income countries with the largest meat-eaters in Australia, consuming around kilograms per person in The average European and North American consumes nearly 80 kilograms and more than kilograms, respectively. However, changes in consumption in high-income countries have been much slower — with most stagnating or even decreasing over the last 50 years.
Consumption trends across Africa are varied; some countries consume as low as 10 kilograms per person, around half of the continental average. Higher-income nations such as South Africa consume between kilograms per person.
One of the strongest determinants of how much meat people eat is how rich they are. This is at least true when we make cross-country comparisons.
In the scatterplot we see the relationship between per capita meat supply on the y-axis and average GDP per capita on the x-axis. What we see is a strong positive relationship: the richer a country is, the more meat the average person typically eats.
Overall, countries tend to shift upwards and to the right: getting richer and eating more meat. What preferences do we have in terms of the types of meat we eat?
Consumption trends vary significantly across the world. In China, pigmeat accounts for around two-thirds of per capita meat consumption. In Argentina, beef and buffalo meat dominates, accounting for more than half of consumption. The visualization details the total number of livestock animals slaughtered for meat in the given year. This is shown across various types of livestock. Here these figures represent the total number slaughtered for meat production which does not include those use primarily for dairy or egg production which are not eventually used for meat.
In , an estimated 69 billion chickens; 1. This is not to be confused with figures above which represent the total number of livestock animals slaughtered or used for meat in any given year. You can find data and research on fish and seafood production and consumption across the world in our entry here. You can find in-depth statistics and research on the environmental impacts of meat and dairy, versus other food products in our article here.
In both examples, the goodness of fit of predicted models was calculated against the input data at the highest administrative level. Figure 2 shows the global distributions of cattle, pigs and chickens and the partial distribution of ducks respectively; created by merging the results from the continental tiles for each species.
These global maps represent the predicted data, first corrected to match the polygon values of the observed data and then to match the FAOSTAT country values in Desert areas and the tropical rain forests of Amazonia and of the Congo Basin have practically no cattle.
The highest concentrations of pigs are found in China and in other Eastern Pacific countries Figure 2b. Pigs are also densely distributed in European countries while only a few countries in Africa e.
Relatively high concentrations are also found in Central America and in Brazil. The distribution of chickens Figure 2c closely follows that of the human population. The highest concentrations are found in eastern China, in Pakistan and India, and in western Europe. In Africa, the countries facing the Gulf of Guinea and Madagascar also have high chicken densities. The densely populated east coast of the United States also shows high numbers whilst chickens are only sparsely distributed in the central and western states.
The heavily populated areas of southern Brazil also show high concentrations of chickens. The distribution of ducks Figure 2d , for those regions for which sub-national statistics were available, adds information to previous national and regional duck mapping efforts [18] , [19].
Ducks are far less common than chickens worldwide although high densities are to be found in South-East Asia and China where duck production is often integrated with rice cropping and fish farming [56] — [57]. Though to a lesser extent, duck densities are also quite high in a few European countries e. Figure 3 zooms in on Thailand and its neighbouring countries and compares the original corrected distribution of poultry a with the new corrected maps of chickens b and ducks c.
Such thematic disaggregation is particularly important in this region because of the abundance of ducks and their important role in farming systems in this part of the world. The figure also illustrates that whilst GLW explicitly excluded the unsuitable areas based on expert opinion, GLW 2 applies a more conservative approach and leaves to the model to predict the livestock distribution.
The two approaches however agree substantially in their mapping of areas with low or zero densities. The finer spatial resolution of the GLW 2 maps approximate cell size of 1 km 2 at the equator also significantly increases the detail of predictions compared to the original version approximate cell size of 5 km 2 at the equator. Figure 4 helps to illustrate this with an example from central Uganda. In this case, the finer pixel size combines with a higher disaggregation of the training data; input data were at administrative levels 1 and 4 respectively in the GLW and GLW 2.
As so very often the case, it is the gathering and processing of primary data — the sub-national statistics on livestock numbers — that is the most arduous and time-consuming part of the process.
This operation is necessarily very labour-intensive and usually requires cases-by-case judgements to be made, in particular regarding the linking of reported statistics to geographical units.
Few advances could be made, therefore, with respect to earlier approaches. The development of GLIMS, however, has greatly reduced the time and effort needed to store, query and prepare the livestock statistics for modelling.
More importantly, the modelling procedure is easily repeated; the suite of R scripts is readily adapted to modelling recurrent updates, different species, different geographical tiles and different technical specifications for statistical modelling and aggregation of results. Automated post-processing activities further allow the rapid production of country-corrected maps for any specified year, global merging of continental tiles, and aggregation to coarser spatial resolutions, as needed.
Computer processing time was primarily a function of the extent of the continental window i. Processing time varied from a minimum of 17 hours for the duck model in Europe to a maximum of hours or 8.
Table 2 summarises the fitting metrics for each tile, species and stratification scheme. It reports both the correlation coefficients and the RMSE between observed and predicted values in the validation data sets. The correlation coefficient is an indication of the precision of the predictions, i. However, even with a nearly perfect correlation, the predicted values can be wrong in absolute terms if, for example, they systematically overestimate the population.
The RMSE, in contrast, is an indicator of the accuracy of the predictions, i. The highest correlation coefficients between observed and predicted values are typically found for the species and tiles for which observed data are at a higher spatial resolution and are evenly distributed within the modelling window. The models for the Asian tile had the best correlation coefficients compared to models for other continents: pigs and ducks 0.
Better accuracy of predictions was generally found for cattle compared to other species with RMSE values as low as 0. RMSE values for cattle and pigs were consistently lower than they were for chickens. Results in Table 2 also indicated, for a given species, that the best stratification differed across the six continental tiles. Composite stratifications had the lowest RMSE values for all species in Asia, but this was the only tile for which the composite prediction performed consistently better than one of the individual stratification schemes.
The biomes stratification performed the best in the North American tile, regardless of the species being modelled, and the EZ25 stratification consistently best in South America. The results of the tests carried out to evaluate the influence of the administrative level of input data on the accuracy of the predictions are presented in Figure 5. In both countries and species tested, the accuracy of the predictions decreased increasing RMSE values as the administrative level of the input data became coarser.
It is notable also that the RMSE values for the cattle predictions in Brazil were lower at all administrative levels than those for the chicken predictions in Thailand. The new GLW 2 livestock density maps described above provide a timely update of the GLW livestock distributions [16] and the enhnaced methods and automated procedures mean that updates will, in future, be more frequent than they have been to date.
The differences in the modelling procedures, the type and resolution of predictor variables and the different unsuitability masks prevent an explicit, quantitative assessment of change between the livestock densities as mapped by the old and new GLW version. However, we can discuss the improvements that have been made over the original GLW, the accuracy of the predictions and areas where further advances could be made.
The first and most obvious improvement is that GLW 2 provides more contemporary data on the distribution of livestock. The original GLW was published in , but many of the reported statistics on which the predictions were based dated back to the 's. Whilst some data in the current set of reported statistics remain relatively old, especially in poor countries where censuses have not been carried out for many years, the median year for which livestock statistics have been obtained is much more recent, with most input data dating being more recent than see file S4.
The second important improvement is in the spatial resolution of the training data, allowing the full potential of the higher spatial resolution 1 km in the predictor variables to be availed. This is clearly shown in Figures 5 , which illustrates how the use of higher resolution training data and predictor variables have allowed the spatial disaggregation of livestock census data at an unprecedented spatial resolution.
Whilst the 1 km resolution may not be required for many applications, such as for example risk or impact assessments at global or continental scales, the higher resolution provides more opportunities for analysis and modelling to be carried out at the scale of individual countries; greatly extending the use of GLW 2. In some countries, the level of detail in the input data is so fine as to bring into question the usefulness of further disaggregating livestock densities to pixel-level.
The reasons for continuing with the modelling, even for small polygons, are four in number. Firstly, even small administrative units are rarely homogeneous in terms of land use and farming so, provided that the covariates are accurate at the finer resolution, the data can still be spatially disaggregated in a meaningful way. Secondly, the training data come in heterogeneous administrative units, which would be otherwise difficult to compare between countries.
Spatial disaggregation to pixel level harmonises the spatial scale of the estimates and thus facilitates comparison. Thirdly, even fine-scale input data often have gaps, caused for example by administrative units where no census data are available or are of restricted access.
The modelled distributions allow prediction of livestock densities to be made in those areas. The fourth reason is possibly the most important. Statistical models based on smaller administrative units will capture more of the underlying environmental variability, and are statistically more robust. A third important improvement in GLW 2 is the estimation of uncertainty around the predicted livestock densities. Whilst the modelling procedure does not provide a full and explicit integration of uncertainty based on a Bayesian framework, the bootstrapping procedure allows coefficients of variation to be calculated around the mean estimates; provides an indicative estimate of the degree of consensus among the different models around the mean estimate.
Finally, the breakdown of species and species groups has also been changed. For example, the previous estimates of poultry densities have been replaced by separate estimates for chickens and ducks. Whilst only cattle, pigs, chicken and ducks have been presented here, GLW 2 is currently being applied to produce distribution maps for other species including goats, sheep and buffaloes, as well as regional maps of camels and equines. This level of species detail is essential for some epidemiological investigations, for example in understanding the geographical distribution of highly pathogenic avian influenza HPAI H5N1, where treating chickens and ducks separately has greatly enhanced our ability to predict risk [29].
These new GLW 2 outputs are already finding applications in diverse fields. The new pig distribution maps are currently being applied to a global analysis of current and future pig production, in which the pig distributions are further disaggregated by production system. The latest version 5 of the GLPS [5] , which is based on land cover data and agro-ecological conditions, maps potential systems rather than actual livestock systems. Integrating these livestock distribution maps with the GLPS will help bring these estimates closer to reality by showing where livestock actually are, rather than where we think conditions are suitable for them to occur.
In combination with the revised agricultural projections to and , recently released by the Agricultural Development and Economic Division ESA of FAO [3] , these new maps are proving important inputs for further efforts to map the demand and supply of animal-source foods [7].
Likewise, updated global maps of livestock distributions will continue to support analyses that quantify greenhouse gas GHG emissions from both the ruminant and monogastric livestock sectors [58]. Finally, the revised maps will find many important epidemiological applications to study disease risk and estimate the impacts of diseases not only in livestock but also for those zoonotic diseases that spill over into the human population.
The poultry distributions have already been incorporated into a number of HPAI H5N1 risk assessments [35] , [59] , and their importance has been highlighted in assessing the risk of spread of the recently emerged H7N9 virus in China and beyond [60] — [61]. Future updates of GLW 2 to include new data as they become available will be greatly facilitated by the integration of the modelling procedure into a fully scripted workflow.
The post-processing is part of this automated workflow and allows a variety of outputs to be derived with relative ease and speed. The reference output presented here is a global map of animal densities at 1 km resolution adjusted at the country level to match FAOSTAT totals. However, outputs expressed in absolute numbers rather than densities , aggregated to different spatial resolutions 5 km, 10 km or 20 km , or matching different FAOSTAT country totals can also be derived.
The accuracy of the predictions varied significantly according to species and the level of detail in the input training data, and were comparable to those found by Van Boeckel et al. The administrative level of the training data had a strong influence on model accuracy; with greater accuracy observed with higher administrative levels of training data for both cattle in Brazil and chickens in Thailand Figure 5. The modifiable areal unit problem MAUP , well known to quantitative geographers [62] , has been demonstrated to have varying effects with different levels of aggregation of input data.
Even though the absolute levels of accuracy in prediction for chickens in Thailand and cattle in Brazil were quite different, the fact that we found a positive association between model accuracy and the administrative level at which the training data were used suggests that, wherever possible, data should be sought and collected at the highest possible resolution to train models, even if this results in heterogeneity in administrative levels used to train models, a result comparable to the one found by Van Boeckel et al.
Another observation is that the predictions in Europe and North America for cattle and ducks, whose distributions are conditional on presence of suitable pasture and access to water, were consistently better i. In contrast, in a continent like Africa, where the majority of chicken and pig production is extensive, the accuracies of the models for cattle, pigs and chicken were comparable.
These observations relate to the influence of the intensity of livestock production on the accuracy of the predictions. As production intensifies it becomes increasingly detached from the land resource base for example as feeds are brought in that are grown in completely different places and thus more difficult to predict based on spatial, agro-ecological variables. This effect is particularly marked for chickens and pigs, where the locations of intensive farming units often have more to do with accessibility to markets or to inputs of one sort or another, than to the agro-ecological characteristics of the land that can be quantified through remotely sensed variables.
In support of this, Van Boeckel and colleagues [63] found significantly lower predictability for models of intensive chicken production than for models of extensive chicken production in Thailand. As chicken production becomes intensified in many countries, chicken distributions will become increasingly difficult to predict geographically using largely agro-environmental predictors.
This inverse relationship between predictability and the level of intensification is well illustrated by the case of chickens in Europe the model which performed least well, with an RMSE value of 1. The regression statistics suggest that the environmental predictors can only partially explain the distribution of chickens and that other factors that are not currently included in the regression models, relating to policy and economics, for example, are likely to be relatively more important in determining the distributions of chickens in such settings.
With general trends toward intensified production, the modelling could be improved in the future by incorporating a wider set of anthropogenic, socio-economic and perhaps demand- or trade-related predictor variables. For intensive production systems, the intrinsic variability is also expected to be higher.
A chicken production unit, falling into a single 1 km 2 pixel, may contain up to a million birds. The characteristics of that pixel, in all covariates, may be identical to those of a neighbouring pixel where an equivalent production plant was not installed. However, the difference in numbers will be so high that the stochastic decision to place the plant in one pixel rather than the other equivalent one will increase the variability that the model cannot capture, even with the most pertinent covariates.
A true validation of the predictive accuracy of these models would involve field observation of livestock densities in different pixels and testing those observations against predicted values. However, livestock census data are generally collected and distributed by area administrative units and so validation would have to be done on re-aggregated model predictions.
Moreover, such validation would be extremely costly and time-consuming. Aside from artificially degrading the training data contributing to models, as we have done here Figure 5 , to make an internal validation of the disaggregation efficiency, there are few options for validation.
Geo-Wiki uses Internet crowd-sourcing to validate and modify land cover information. It is a web-based application with Google Earth as a backdrop, over which various global land cover datasets are draped. Discrepancies in land cover assignation among the datasets can be highlighted and values may be either corroborated or changed by users logged onto the wiki. The development of the Livestock Geo-Wiki was proposed by Robinson and colleagues [5] as a way to validate livestock production systems information and it is hoped that some form of validation will also be possible for livestock distributions.
The roles of the Livestock Geo-Wiki for the livestock distribution data described here are, however, in viewing and providing open access to the data. An important objective of the Livestock Geo-Wiki is to disseminate the data. Whilst this has provided a valuable information resource it is not interactive so the opportunities for users to provide feedback are limited.
The Livestock Geo-Wiki has several advantages over more static ways of disseminating data, including, for example: a providing a central repository where many different aspects of livestock information may be explored in a highly interactive way; b publicising and disseminating open-access livestock sector data at a range of spatial scales; c using crowd-sourcing approaches to validate and improve livestock data; and d providing innovative data visualisation and analysis tools that will facilitate the investigation of dependencies among the data sets and address specific requirements of diverse groups of users.
The improvements to global livestock distribution data, presented here, have been motivated by the pressing need for higher resolution and more contemporary outputs than provided by the original GLW livestock distributions.
Although the methodology has been considerably revised in many aspects, it remains similar to that developed by Wint and Robinson for GLW [16] , involving the use of several stratified multiple regressions linking observed livestock densities to environmental data.
New machine-learning methods such as Boosted Regression Trees, or Random Forests have recently demonstrated their increased predictive capacities over more standard statistical methods for species distribution modelling [66].
These methods can better account for non-monotonic relationships between dependent and predictor variables, and can better incorporate the effects of interactions among predictors.
Currently, the comparative strength of the stratified regressions lies in the relatively fast fitting of models, and their rapid application to large data sets. However, with the advent of parallelised computing, the use of such machine-learning techniques for modelling livestock distributions is becoming a realistic prospect. These approaches should be investigated in relation to their potential superiority in terms of predictive capacity. In addition, alongside the current set of predictor variables, which are principally environmental in nature, incorporating more socio-economic and anthropogenic predictor variables may confer considerable improvements to models for which the livestock distributions are dominated by more intensified modes of production.
Considerable progress has been made in advancing the spatial modelling of livestock distributions, and there are clearly areas where further improvements can be made. No matter how good the predictive model though, it will always be limited by the quality of the training data on which is based.
With this in mind, efforts must be made to identify areas where reported statistics on livestock distributions are out-dated, of low spatial resolution or of doubtful quality, and steps must be taken to address these shortfalls.
Detailed information on MODIS-derived predictor variables source, derived Fourier variables, image values and rescaling if applicable. Islamic countries in relation to modelling the distributions of pigs. Countries emboldened in red are those flagged with zero values in modelling the distribution of pigs. The burgers, steaks and sausages served up in most wealthier countries tend to be a lot bigger than the recommended 70g a day.
The livestock economy is particularly important for poor rural populations in low- and middle-income countries. Some of these involve the cultivation of animal cells in labs — growing real meat in a petri dish rather than using an animal. Another approach is the engineering of plant- or fungi-based meat substitutes, to give them the taste and texture of beef, pork or chicken.
And there are attempts to make insects — already eaten in parts of Asia and Africa — a more popular choice on menus worldwide. For millions of people, eating animals is a way of life — one of the cultural cornerstones of their domestic and social lives.
The views expressed in this article are those of the author alone and not the World Economic Forum. In-depth knowledge of the supply chain in food production can help us reduce carbon emissions.
Companies must work together at each stage, say experts. Climate change-related factors such as water stress, increased temperatures and carbon dioxide can all reduce the quality of coffee, according to research. I accept. Take action on UpLink. Forum in focus. Read more about this project. Explore context. Explore the latest strategic trends, research and analysis. Billions of animals are slaughtered every year.
But those figures are dwarfed by the number of animals we eat. We eat more meat per person than ever. Meat production costs the Earth. Have you read?
0コメント