Rangeland Resources Research Unit, United States Department of Agriculture-Agricultural Research Service, 1701 Centre Avenue, Fort Collins, CO 80525, USA.
Sensors (Basel). 2013 Mar 15;13(3):3711-23. doi: 10.3390/s130303711.
Advances in global positioning system (GPS) technology have dramatically enhanced the ability to track and study distributions of free-ranging livestock. Understanding factors controlling the distribution of free-ranging livestock requires the ability to assess when and where they are foraging. For four years (2008-2011), we periodically collected GPS and activity sensor data together with direct observations of collared cattle grazing semiarid rangeland in eastern Colorado. From these data, we developed classification tree models that allowed us to discriminate between grazing and non-grazing activities. We evaluated: (1) which activity sensor measurements from the GPS collars were most valuable in predicting cattle foraging behavior, (2) the accuracy of binary (grazing, non-grazing) activity models vs. models with multiple activity categories (grazing, resting, traveling, mixed), and (3) the accuracy of models that are robust across years vs. models specific to a given year. A binary classification tree correctly removed 86.5% of the non-grazing locations, while correctly retaining 87.8% of the locations where the animal was grazing, for an overall misclassification rate of 12.9%. A classification tree that separated activity into four different categories yielded a greater misclassification rate of 16.0%. Distance travelled in a 5 minute interval and the proportion of the interval with the sensor indicating a head down position were the two most important variables predicting grazing activity. Fitting annual models of cattle foraging activity did not improve model accuracy compared to a single model based on all four years combined. This suggests that increased sample size was more valuable than accounting for interannual variation in foraging behavior associated with variation in forage production. Our models differ from previous assessments in semiarid rangeland of Israel and mesic pastures in the United States in terms of the value of different activity sensor measurements for identifying grazing activity, suggesting that the use of GPS collars to classify cattle grazing behavior will require calibrations specific to the environment and vegetation being studied.
全球定位系统（GPS）技术的进步极大地提高了跟踪和研究自由放养牲畜分布的能力。了解控制自由放养牲畜分布的因素需要能够评估它们何时何地觅食。在四年（2008-2011 年）期间，我们定期收集 GPS 和活动传感器数据，并结合对科罗拉多州东部半干旱牧场放牧牛的直接观察。根据这些数据，我们开发了分类树模型，使我们能够区分放牧和非放牧活动。我们评估了：（1）GPS 项圈中的哪些活动传感器测量值最有助于预测牛的觅食行为，（2）二进制（放牧、非放牧）活动模型与具有多个活动类别的模型（放牧、休息、移动、混合）的准确性，以及（3）在多年间稳健的模型与特定于特定年份的模型的准确性。二叉分类树正确地去除了 86.5%的非放牧地点，同时正确地保留了 87.8%的动物正在放牧的地点，总体错误分类率为 12.9%。将活动分为四个不同类别的分类树产生了更高的错误分类率 16.0%。在 5 分钟间隔内行驶的距离和传感器指示低头位置的间隔比例是预测放牧活动的两个最重要的变量。拟合牛觅食活动的年度模型并没有比基于所有四年的单一模型提高模型准确性。这表明，增加样本量比解释与草料产量变化相关的觅食行为的年际变化更有价值。我们的模型在半干旱牧场以色列和湿润牧场美国的放牧活动方面与以前的评估有所不同，这体现在不同活动传感器测量值识别放牧活动的价值上，这表明使用 GPS 项圈来分类牛的放牧行为将需要针对所研究的环境和植被进行校准。
