Cleveland Metroparks has deployed a permanent array of over 200 wildlife cameras on plots to monitor animal communities and activity throughout our 23,000 acre park system. These plots were chosen using stratified random methods to extend inference to the whole park system. Our main questions were whether such an array could be used to estimate biodiversity, monitor species of conservation concern, and estimate activity of several species of management interest. This array has generated over 12 million images in 5 years. When we began this project, no other project had attempted to use such a large array of permanently placed cameras. We developed an open-source machine learning (ML) classifier, based on a convolutional neural network implemented using Tensorflow on an Amazon Web Services platform. This model was trained using up to 2000 images per class for 32 classes, where most classes were single species and several included multiple, taxonomically similar species. This training set was generated from images taken in the first 9 months of 2016. We compared estimates of biodiversity and activity between the human and model classified data.
Results/Conclusions
Using the ML model we have classified over 3000 images from the training period and over 1 million images from 4 months in 2018. We have developed processes and software for manipulating large volumes of images. Model accuracy was greater than 80% for species of interest such as coyote and white-tailed deer and 99% for turkey. Estimates of biodiversity and activity indices generated by the ML model are broadly comparable to those generated from human tagged images. We present these comparisons for 5 species of interest. We discuss the usefulness of such a large array of permanently mounted wildlife cameras for use in managing wildlife in a large urban park system. We will present examples of products of the model that can be used by managers, such as heat maps of species' activity, that can be compared across time.