Using the XGBoost algorithm to classify neck and leg activity sensor data using on-farm health recordings for locomotor-associated diseases

• Published in: Computers and electronics in agriculture Vol. 173; p. 105404
• Main Authors: Gertz, M., Große-Butenuth, K., Junge, W., Maassen-Francke, B., Renner, C., Sparenberg, H., Krieter, J.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.06.2020; Elsevier BV
• Subjects: Algorithms; Automation; Classification; Computing time; Motion sensors; Segmentation; Sensors
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105404

•Locomotor-associated diseases could be successfully classified (86% AUROC & 81% F-Measure).•Computational load could be reduced by approximately two-thirds using feature pre-selection.•Different feature types and window-lengths were considered.•XGBoost is a capable and easy to use method for classification of motion-sensor data in cattle. The objective of this paper was to analyse the feasibility of an automated detection system for locomotor-related diseases under strictly practical conditions. For this purpose, motion sensor data and on-farm health recordings of locomotor-, respiratory- and udder-health issues of 397 cows on a commercial farm in Germany were recorded, employing two sensors at the leg and neck simultaneously. Sensor Data and health recordings were evaluated with the XGBoost algorithm to predict “sick” and “healthy” cows. Classification examples were constructed by dividing sensor data based on eight different window lengths and by applying four different features to produce the resulting segments. Results suggest that sensor information and health recordings can be utilised successfully to learn sickness-behaviour patterns, which enable a classification of data excerpts to “sick” and “healthy” by achieving 86% (±2%) AUROC, 81% (±2%) F-Measure as well as relatively balanced specificity (78%) and sensitivity (81%) levels when using all variables and features available. Results further indicate that sensible feature selection can reduce computing time greatly by only minor losses in classification performance. XGBoost classification seems to be a powerful, easy-to-use and efficient method for the identification of sickness behaviour under practical conditions. Nevertheless, the results also indicate that more research is necessary into the choice of features, the ideal window length(s) or segmentation strategy, interaction of feature variants and variable preselection. Overall, the applied approach indicates a substantial potential for the development of an automated detection tool in the future, capable of an ongoing evaluation of locomotor-associated diseases.