Robust ecological analysis of camera trap data labelled by a machine learning model

Abstract

1. Ecological data are collected over vast geographic areas using digital sensors such as camera traps and bioacoustic recorders. Camera traps have become the standard method for surveying many terrestrial mammals and birds, but camera trap arrays often generate millions of images that are time-consuming to label. This causes significant latency between data collection and subsequent inference, which impedes conservation at a time of ecological crisis. Machine learning algorithms have been developed to improve the speed of labelling camera trap data, but it is uncertain how the outputs of these models can be used in ecological analyses without secondary validation by a human.2. Here, we present our approach to developing, testing and applying a machine learning model to camera trap data for the purpose of achieving fully automated ecological analyses. As a case-study, we built a model to classify 26 Central African forest mammal and bird species (or groups). The model generalizes to new spatially and temporally independent data (n = 227 camera stations, n = 23,868 images), and outperforms humans in several respects (e.g. detecting 'invisible' animals). We demonstrate how ecologists can evaluate a machine learning model's precision and accuracy in an ecological context by comparing species richness, activity patternsThis is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.

Methods for wildlife monitoring in tropical forests: Comparing human observations, camera traps, and passive acoustic sensors

“…Annotation and management of such volumes can be challenging for monitoring projects (Glover-Kapfer et al, 2019), despite the availability of various platforms for data management (Young et al, 2018). Image annotation by automated classification is developing rapidly (Glover-Kapfer et al, 2019;Whytock et al, 2021;Willi et al, 2019) and is increasingly being integrated in data management platforms (Ahumada et al, 2020) and desktop apps (Falzon et al, 2020), requiring gradually less technical expertise and improving access for mainstream use (Aodha et al, 2014). Algorithms can annotate images with increasing accuracy to species or genus level, or filter out empty images (Wei et al, 2020), which can drastically reduce the workload (Norouzzadeh et al, 2018;Tabak et al, 2019).…”

Estimating the movements of terrestrial animal populations using broad-scale occurrence data

“…Thus, one challenge with acoustic and camera trap data is the combination of technical software and skill needed to identify and isolate the correct sounds or images for analysis [ 97 ], but also the human time that is often needed to manually validate portions of the data for accuracy [ 96 , 98 , 99 ]. Sparse arrays of acoustic or camera monitors may be useful for confirming a species’ occupancy, or for estimating animal activity patterns, abundance, or species diversity in an area, especially when robust methods for confirming species presence have been developed [ 100 , 101 ], but much larger and denser arrays would be needed to infer population movement through or within an area [ 46 ]. Differences in camera trap survey designs, including baited versus unbaited stations, have been found to have significant consequences for occurrence frequency and detection rates [ 102 , 103 ].…”

“…Variable sampling effort over space is often accounted for using covariates that are suspected to correlate with sampling effort, such as distance to the nearest urban center (e.g., [ 137 ]) or distance to road [ 68 ]. In some cases effort covariates do not exist (e.g., iNaturalist [ 101 ]) and researchers must instead control for variable sampling effort with other methods, such as using the number of non-target species detections as a way to estimate change in effort across space and time [ 138 ]. Another common approach used when fitting species distribution models is to sample background locations in a way that attempts to mimic sampling biases in the occurrence data [ 139 , 140 ].…”

Domain-Aware Neural Architecture Search for Classifying Animals in Camera Trap Images

“…Edge computing [ 16 ] was ideally developed for such cases, i.e., intensive computation centralized at a data center can be split and localized by edge devices near camera traps [ 17 , 18 ]. Thus, fundamental processing steps such as removing images without animals [ 6 , 7 , 12 , 19 , 20 , 21 ] and classifying images with animals [ 9 , 10 , 11 , 12 , 13 ] can be automatically conducted on edge devices. However, edge devices are not only heterogeneous [ 22 ] but also resource constrained [ 23 ].…”