PSXI-7 Sensing of Defecation and Urination Events Through Carbon Dioxide Monitoring

Abstract

Abstract The continued success of livestock production in grazing systems depends on the ability to manage lands sustainably. Non-point-source emissions, particularly defecation/urination events (DUE), are a major cause of soil degradation, runoff into water sources, and nutrient management issues. Monitoring of these emissions from livestock through wearable sensing could support more precision management, but is complicated by the need for individualized sensors, variable signal strength across topographies, and unreliable data transmission in remote locations. The objective of this study was to explore opportunities to use CO2 sensing as an indicator of DUE from cattle on pasture using data obtained from an open-source sensor. The use of tail sensors was implemented to monitor the DUE of horses at various times of the day, ranging from 8 am to 5 pm. The tail sensors were developed using a neoprene base attached to velcro straps that were placed around the tailhead. The sensors themselves were equipped with a SparkFun ESP32 Thing Plus microprocessor connected to a SparkFun CO2 sensor and a lithium ion rechargeable battery that was charged via micro USB (SparkFun Electronics, Niwot, CO). Sensors were placed on 4 horses for 6 hours each collection period and data collection occurred over 9 days. Horses were observed in real time and DUE were recorded at the time of occurrence. Data were collected from the sensors via a Dragino LoRa gateway and a mobile hotspot for wireless connection. Data analysis relied on visual data exploration and targeted assessment of associations using random forest regression, support vector machines, and extreme gradient boosting. The visual data exploration revealed that DUE occurred in much lower frequency than non-events, and so random oversampling and undersampling were used for each classification approach to attempt to improve accuracy. Although CO2 peaks often corresponded to DUE, they also frequently occurred in the absence of DUE. All classification algorithms showed poor accuracies (0.50 to 0.51), which were only marginally improved by over- (< 0.51) and undersampling (< 0.69). This preliminary assessment revealed considerable noise in sensing CO2 emissions in production settings, which may preclude usefulness in DUE sensing.