Development of a smart livestock farming tool for identifying animal growth using artificial intelligence

Development of a smart livestock farming tool for identifying animal growth using artificial intelligence

Affordable tools with the ability to continuously monitor the growth rate of livestock animals are highly sought after by the livestock industries. This demand is driven by the potential for these tools to assist in improving animal welfare and production efficiency. In a rapidly growing population the demand for meat is escalating, especially in Asia, where the middle class is currently expanding. Meanwhile in the western world there is growing consumer concern surrounding animal husbandry, with certain organisations labelling some of the current husbandry practices cruel or sub-standard. The environmental impacts of livestock farming are also increasingly becoming scrutinised, pressuring researchers to find new methods to increase the efficiency of livestock nutrition, and improve health (disease prevention), reproductive and waste management practices. At the centre of these problems is the ever-changing individual animal as it continuously adapts to its surrounding environment and available resources.

Livestock growth is a fundamental measure which can be used for diagnostic purposes in these areas, therefore the main objective of this study was to develop a system to automatically determine the growth of individual and groups of livestock animals (pigs) using welfare friendly and non-invasive methods. A machine vision system was selected to undertake this weight estimation task, whereby pigs’ body measurements are extracted from images and used to estimate their weight without physical interference.

Reviews prompted the development of a methodology to determine the weight-estimation equations as a function of not just the animals’ body measurements but also their pose. Subsequently equations were generated from shapes that conformed closely to a specified reference template shape. Thus, to enhance precision during weight estimation the template shape was directly linked to the equation and pose validation aspects of the system. Filters were developed to provide recognition via the confirmation of the characteristic template shape and known body measurement and weight relationships. The shape filter ensured that 94% of weight estimates that passed through to output were within ±5 kg of the actual weight of the pig. Using the shape and limit filter in unison ensured that greater than 97% of the samples which passed had an weight estimate within ±5 kg of the actual weight of the pigs and 68% of the total number of samples were within ±2 kg. Statistical modelling was used to determine the importance of different body measurements in estimating weight. Subsequently a multivariate linear weight estimation equation was created to estimate pigs’ weight using a stepwise selection of variables. The multivariate linear equation estimated 2% more sample weights within ±2 kg error and 3% less sample weights greater than ±5 kg error than the closest non-linear equation. Software was written to automatically recognise pigs inside the field of view (FOV) of the camera and to extract 16 body measurements from the pigs’ body contours. Height was manually recorded from the back of a sample of pigs to determine its strength in weight estimation. Including the pig’s height in the weight-estimation equation did improve predictive performance with a 7.34 % improvement in the number of samples estimated within ±2 kg of the pigs’ actual weight compared to a multivariate equation without the height parameter. Although, this improvement was not significant enough to justify the additional practical development required to collect the height information automatically during the weight estimation process.

Both off-line simulation and on-farm experiments were undertaken using data collected from commercial facilities. During an off-line simulation, the shape and dimension filters were applied across a dataset containing over 20,000 frame samples of over 500 pigs. Gut fill was used as a guide to determine a practical error margin for measuring the weight of individual pigs across the course of a day. The machine vision system was found to operate within an acceptable error margin of 50 % of the gut fill according to the equation and average shape template used during off-line simulations. As on average pigs in the weight-range of 45 to 115 kg had their live body weight estimated to within 3.16 % and 2.20 % of their actual live body weight, respectively. For pigs less than 45 kg in weight the piGUI system operated, on average, to within 67% of the weight attributed to gut fill (between ±1.07 and ±1.49 kg error). During off-line simulations, the percentage mean-relative error obtained by the piGUI system was between 5.1 and 3.7% for pigs in the weaner to grower weight range (15 to 45 kg) and less than or equal to 2.5% for grower finisher pigs between 45 and 115 kg. Thus, on average, the system was able to estimate the pig’s body mass with practical precision.

The system labelled ‘piGUI’ was installed in pens at commercial facilities which housed pigs in group-sizes of between 10 and 160 pigs. During testing, the system determined the average weight of groups of pigs on a daily basis, tracking the group’s growth rate. In some trials, the pig’s weights were also estimated along with the weight deviation of the group. During a 22 day trial period the system estimated the average weight of a group of finisher pigs within 2.1%, on the seven days when the actual group weight was recorded from an electronic scale. No information was passed between successive days by the system.

The diagnostic power of the piGUI system was also tested on-farm. A deflection away from the standard growth curve was recorded during two successive batches of grower pigs after reaching weights greater than ~45 kg. These growth deflections were believed to be caused by stress related directly or indirectly to temperature, as the summer temperatures reached over 38°C during these batches. The level of animal activity recorded by the system, the temperatures leading up to the deflection in growth and figures reported in literature support this theory.

The piGUI system was also tested to see whether it could estimate the weight of sows in their early stages of pregnancy and whether it could detect changes in the body measurements of individual sows before and after giving birth. A group of eleven sows between day 71 and day 82 of pregnancy had their group weight estimated to within 0.1 kg of their actual group weight. Eighty-two percent of their individual weights were estimated within a practical range of ±5 kg of their actual weight. The metric body measurements of two Large White × Landrace sows were also recovered by the vision system before and after giving birth. The widths and lengths of the sows’ recorded by the vision system were consistent with those found in literature. Indicating that the device may be used to monitor sow weight and body morphology in future.

The developed device was also tested at various locations within the pen environment. Radio Frequency Identification (RFID) was integrated into the system to determine whether bias in group estimates could occur as a result of the sampling region observed within the pen. A layout bias was discovered, caused by certain pigs visiting the FOV (containing the feeder) more frequently or for longer durations than others. Subsequently, feeding behaviour was determined using the RFID information collected and demand for the feeder was calculated for the pigs individually and as a group. The number of social interactions between pigs at the feeder was also determined, thus providing a method to identify social interaction and potentially the competitive nature of pigs automatically.

A comparative study was undertaken between a commercial system ‘System-A’ and the piGUI system. System-A failed to correctly estimate the group average weight of the finisher pigs in the trials. It was apparent that necessary conversions were not taking place within System-A’s software to normalise the extracted body measurements to suit weight-estimation equation coefficients. It was found that, System-A’s growth data would require a multiplication by a scalar factor to adjust the growth data to valid weight ranges. Code within the piGUI software performed the necessary conversions automatically during initialisation and was not burdened by this limitation. The piGUI system estimated the group average weight to within 2.1% on each of the seven days when the actual weight of the pigs were determined using the electronic scale. On these days, System-A reported group average weight estimates in excess of 16 kg error of the actual group average weight. It was clear that the distribution of weight data recorded daily by the piGUI system was far more concentrated around a mean estimate value than system-A.

The results of this PhD study demonstrate that the average weight of groups of pigs can be calculated with sufficient practical accuracy. The precision achieved during this study was better than reported in the literature and the precision of the system was also favourable compared to a commercially available system. Therefore the developed system can be used for practical purposes on commercial farms to determine the average weight and growth of groups of grower-finisher pigs.