The recovery of cattle herd spatial location using animal movements
-
1
University of São Paulo, Brazil
-
2
Roslin Institute, University of Edinburgh, United Kingdom
In order to build surveillance and control programs to prevent livestock epidemics a knowledge of the farming system itself is fundamental. On top of characteristics such as the rearing practices in place and the network of animal movements, farm spatial locations are also crucial information to effectively design such programs, as spatial proximity is a useful proxy for many disease risks. A lack of spatial information could therefore compromise the best evaluation of a disease’s potential to spread, especially for livestock diseases where spread has strong spatial characteristics, either through direct transmission or because of common risk factors such as wildlife reservoirs or insect vectors (e.g. Foot-and-mouth disease, bovine tuberculosis, blue-tongue virus, African swine fever, avian influenza). Moreover, it could compromise epidemiological studies that evaluate risk factors associated with the topology, and it could slow down the decision-makers response in the case a new epidemic outbreak. Therefore, methods to compensate for the lack of farm spatial information are very important.
The objective of this study was to provide a methodology to infer farm position in space based upon animal trading exchanges (e.g. contacts) to or from other farms. This methodology was developed and applied to the cattle rearing system of Espirito Santo (ES) State (Brazil). The main information to identify their spatial location within the State are coordinates (latitude and longitude) and municipality (e.g. county). There were 45’419 farm registered in ES in 2017. However, only 7’995 of these had coordinates recorded with enough quality to be used. The municipality was available for all of them. The main hypothesis was that the distance between a target farm and its trade partners can be a good predictor to find its possible location within the specified municipality. For this purpose we used farms which had incoming or outgoing cattle movements in 2016 and 2017. We calculated the geodesic distances of contacts between farms with known locations, and then we fitted the empirical data to a lognormal distribution.
To predict the location of a target farm in its municipality, we divided that municipality into areas of 1km2 and calculated the geodesic distance (km) of each one of those to the trade partners of the target farm. Thus, each area had a set of distances. Those distances were fitted with the log-normal distribution (calculated previously) to obtain the probabilities of the events. Then, they were transformed to a logarithmic scale and summed in each area to have a total weight by area. We considered that the areas with higher total weight were the ones where the target farm was more likely to be located. To assess the locations prediction model accuracy we calculated the distribution of distances from the real to the predicted farms location. As a second accuracy measure, for each location farm prediction we assigned a ranking value to all municipality cells (1 km2 area each) based on the model result, and we compared the ranking of the cell with where the farm was really located. By doing so we could estimate the percentage of municipality area needed to find the true location of each farm. Then, we fitted an empirical cumulative distribution in both of these measures.
To test if our method was better than a random model, we sampled 100 random cells in the municipality for each farm predicted and calculated the distribution of median distances from the real to the random farm location. Then, we fitted an empirical cumulative distribution and compared with the results obtained with our method.
The methodology was tested on 3’393 farm with known coordinates. Our results showed that 50% of the farms were located at less than 6 km radius from the areas predicted, but 9 km less than using a random sample in the municipalities (Figure 1A). Also, it was possible to find 50% of the true farms location with less than 11% of the municipality’ area (Figure 1B).
In order to understand which variable mostly affected the model ability to predict farm position, we built a logistic regression model where the farms’ position outcome (correct or not) was tested against municipality, spatial and network variables. To do this, we chose arbitrarily the 3% of areas with highest total weight to evaluate the quality of the predictions (27% correctly farms’ locations within the municipalities). Then, the dataset was divided in two parts, one dataset to build the logistic model (training dataset) and a second dataset to validate the model (validation datasets), with 70% and 30% of the information respectively. We calculated the curve ROC to evaluate the performance of the model and chose a probability cutoff in that curve to have specificity values ≥ 0.8. In addition, we created 100 random sub-samples, reordering the nodes in the training and validation datasets, running the regression model each time and calculating the sensitivity and specificity of the model with the validation datasets with the same threshold used previously. The result of regression model applied to predict the outcomes in the validation dataset had a median sensitivity of 0.39 (min 0.30 and max 0.47) and a specificity of 0.78 (min 0.74 and 0.85) after the 100 replications.
In conclusion, the locations within a municipality which minimized the average geodesic distance to known contacts was found to be a significantly better predictor of farm location than random locations within the municipalities. However, other variables need to be considered in the methodology in order to improve farms’ location prediction, as we could not predict all the farms in the region. Also, we identified some variables (spatial distribution of the partners around the predicted area, distance between the partners, distance between the municipalities, proportion of partners from the same municipality, distance between the predicted area and the limit of the municipality, and density and number of farms in the municipality) that would help us to understand the farms’ position outcome with our method. So long as the farms with missing locations have similar characteristics to the farms with recorded locations, it should be possible to use this methodology to better identify locations with some robustness.
Acknowledgements
This work was supported by FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo), grant 2018/17762-4, and the BBSRC Institute Strategic Programme, grant BB/P013740/1, for Jason Onell Ardila Galvis.
Keywords:
missing coordinates,
Network analysis,
spatial analysis,
data quality,
Veterinary Epidemiology
Conference:
GeoVet 2019. Novel spatio-temporal approaches in the era of Big Data, Davis, United States, 8 Oct - 10 Oct, 2019.
Presentation Type:
Regular oral presentation
Topic:
Spatial-explicit or spatio-temporal network analysis
Citation:
Ardila Galvis
JO,
Rossi
G,
Cardenas
NC,
Grisi-Filho
JH and
Kao
RR
(2019). The recovery of cattle herd spatial location using animal movements.
Front. Vet. Sci.
Conference Abstract:
GeoVet 2019. Novel spatio-temporal approaches in the era of Big Data.
doi: 10.3389/conf.fvets.2019.05.00037
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
31 May 2019;
Published Online:
27 Sep 2019.
*
Correspondence:
DVM. Jason O Ardila Galvis, University of São Paulo, São Paulo, Brazil, jason.ardila@usp.br