Monitoring diet with automated microhistology

Abstract

AbstractMicroscopic analysis of indigestible fragments (microhistology) has long been used to monitor diets of wildlife, but the method is slow and difficult to replicate. We automated microhistology analysis by training a computer vision network to quantify the percent slide cover of identified fragments. We compared automated and manual analyses for 3 reference mixes with known proportions of plants and for samples of feces collected during the fall of 2016 and 2017 from white‐tailed deer (Odocoileus virginianus) and goats (Capra aegagrus) with access to the same plants in the wild at the Sonora Research Station, Edward's Plateau, Texas. Consistent size and shape of fragments for grass (Bothriochloa ischaemum) and conifer (Juniperus ashei) were associated with absolute errors of 6 to 16% for both methods. Irregular fragments of leaves and acorns of oak (Quercus virginiana) were underestimated by both methods. Small star‐shaped trichomes of a forb (Croton fruticulosus) were underestimated by computer analysis but overestimated by manual analysis. Fecal fragments from deer were more diverse in plant species and more variable in thickness and shape than those of goats. Computer analysis estimated conifer content of the goat diet within 4% of the manual estimate whereas those estimates for deer were up to 35% greater than the manual estimate. Automated analysis of digital images can increase the rigor of microhistology by expanding the size of sample sets and facilitating the routine use of calibrated references to assess errors in estimates of diet.