Unmanned Aerial Vehicle–Based Remote Sensing of Cattle Dung: Detection, Classification, and Spatial Analysis of Distribution

Abstract

Documenting the distribution of cattle dung across grazed pastures is an important part of understanding nutrient cycling processes in grasslands. However, investigation of distributions at adequate spatial scales and over extended time periods is hindered by the lack of a time- and cost-efficient method for documenting and monitoring dung pat locations. To address this research challenge, an unmanned aerial vehicle and multispectral sensor were used to identify and classify dung pats. Imagery was collected on 12 flights over a subirrigated meadow in the Nebraska Sandhills, in which two different grazing strategies were being evaluated: an ultrahigh stocking density and a low stocking density. The images were classified using supervised classification with a support vector machine algorithm, and post-classification accuracy was assessed using a confusion matrix. In addition, Ripley’s K was used to identify high-density dung areas at varying densities and spatial extents. The classification had an overall accuracy of 82.6% and a Kappa coefficient of 0.71. The user’s accuracy of dung classification was higher (0.91) than the producer’s (0.73). The majority of classification errors were related to the misclassification of dung as vegetation, often in spectrally complex areas where shadowing affected the ability of the classifier to correctly identify dung. Classification accuracy declined precipitously after dung reached 10-14 d of age, both because of the change in spectral reflectance due to drying and because of the regrowth of vegetation. The density-based cluster analysis found no clustering in the low stocking density treatment; dung in the ultra-high stocking density treatment was most frequently found to be clustered near water sources, in corners, and near supplement feeders. This approach to dung identification, mapping, and spatial cluster analysis is a promising alternative to existing methods and deserves further exploration at additional spatial scales and in diverse ecological settings using current technologies.