A framework for assessing reliability of observer annotations of aerial wildlife imagery, with insights for deep learning applications

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(1), P. e0316832 - e0316832

Published: Jan. 15, 2025

There is growing interest in using deep learning models to automate wildlife detection aerial imaging surveys increase efficiency, but human-generated annotations remain necessary for model training. However, even skilled observers may diverge interpreting imagery of complex environments, which result downstream instability models. In this study, we present a framework assessing annotation reliability by calculating agreement metrics individual against an aggregated set generated clustering multiple observers’ observations and selecting the mode classification. We also examined how image attributes like spatial resolution texture influence observer agreement. To demonstrate framework, analyzed expert volunteer twelve drone images migratory waterfowl New Mexico. Neither group reliably identified duck species: experts showed low (43–56%) several common species, volunteers opted out task. When simplified into broad morphological categories, there was high cranes (99% among experts, 95% volunteers) ducks (93% 92% volunteers), though notably lower classifying geese (75%) than (94%). The sets from two groups were similar: count birds across all 91% count, with no statistically significant difference per (t = 1.27, df 338, p 0.20). Bird locations matched 81% between classifications 99.4%. Tiling reduce search area maintaining constant scale keep size differences classes consistent Although our sample limited, these findings indicate potential taxonomic limitations show that, aggregate, can produce data comparable experts’. This assist other practitioners evaluating their input

Language: Английский

Chimpanzees (Pan troglodytes) Indicate Mammalian Abundance Across Broad Spatial Scales

Ecology and Evolution, Journal Year: 2025, Volume and Issue: 15(3)

Published: March 1, 2025

ABSTRACT Ongoing ecosystem change and biodiversity decline across the Afrotropics call for tools to monitor state of or elements extensive spatial temporal scales. We assessed relationships in co‐occurrence patterns between great apes other medium large‐bodied mammals evaluate whether ape abundance serves as a proxy mammal diversity broad used camera trap footage recorded at 22 research sites, each known harbor population chimpanzees, some additionally gorillas, 12 sub‐Saharan African countries. From ~350,000 1‐min videos 2010 2016, we estimated mammalian community metrics, including species richness, Shannon diversity, mean animal mass. then fitted Bayesian Regression Models assess potential detection rates (as abundance) these metrics. included site‐level protection status, human footprint, precipitation variance control variables. found that species, well mass were largely positive. In contrast, rate richness less clear differed according site impact context. no association diversity. Our findings suggest chimpanzees hold indicators specific communities, especially population‐level composition‐related characteristics. Declines chimpanzee populations may indicate associated declines sympatric highlight need improved conservation interventions.Changes likely precede extirpation mammals.

Language: Английский

A protocol for error prevention and quality control in camera trap datasets

Journal of Applied Ecology, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 14, 2025

Abstract Camera traps are a mainstream methodology in applied ecology, but surprisingly there no widely accepted protocols to ensure the quality of data obtained from these devices. We reviewed sample 147 articles recent camera‐trapping literature and found that only 4.8% report measure control. propose framework process media files camera minimises errors by adopting series systematic procedures. Before classification, focus is on detecting malfunctions, correcting storage programming establishing clear exclusion criteria. Classification can follow different approaches, including single or double human‐eye review, which be supported artificial intelligence. The protocol followed control procedures enable users determine whether dataset meets standards ready analysed, if further revision needed. Synthesis applications : proposed introduces as key component trap processing, thus reducing error rates making reporting more transparent. These principles also apply other methods, such autonomous sound‐recording units. suggest formal procedures, ecology will able capitalise many advantages brought new technologies processing tools.

Language: Английский