Data from: Supplementing human observation with artificial intelligence impacts demographic estimates for a Critically Endangered lizard.
Data files
Jan 30, 2025 version files 67.89 KB
-
AI_SupplementedEncounterHistory_Collapsed.inp
7.20 KB
-
AI_SupplementedEncounterHistory.inp
24.12 KB
-
HumanOnlyEncounterHistory_Collapsed.inp
7.79 KB
-
HumanOnlyEncounterHistory.inp
26.13 KB
-
README.md
2.66 KB
Abstract
Species conservation relies heavily on population estimates derived from capture-recapture analyses, which are liable to produce biased results if individual animals are incorrectly identified. Captive and known-animal studies have shown that supplementing human observation with artificial intelligence (AI) has the potential to reduce these errors. However, no study has directly quantified the relationship between using AI for individual identification and the demographic estimates it produces for a threatened population in situ.
We compared the demographic estimates produced by capture-recapture analyses of two distinct encounter histories constructed from the same survey data; one produced using individual identifications made by human observers alone (the 'Human-Only data set'), and one produced using AI software to aid individual identification (the AI-Supplemented data set'). This approach enabled us to address two key questions: (i) does the use of artificial intelligence software for individual identification influence demographic estimates for an in-situ conservation programme? (ii) how has the population of our case study species, the Critically Endangered Kapitia skink, responded following an extreme weather event, cyclone Fehi?
We found that without AI, human observers appeared prone to make reclassification or 'splitting' errors, in which a recaptured animal was wrongly assigned as a new individual. Analysis of the AI-supplemented data set consistently produced lower estimates of population abundance over time, relative to the same analysis of the Human-Only data set. This provides new evidence that wild species monitoring efforts may be prone to underestimating the extinction risk of populations if they are dependent on individual identification methodologies with high potential for human errors.
Our case study species, the Kapitia skink, demonstrated a positive population trend in the period following cyclone Fehi. Whilst promising, conservation intervention is recommended to address persistent threats.
Practical Implication: Supplementing human observation with AI software for individual identification could mitigate errors leading to the underestimation of extinction risk for endangered species. We encourage further development of AI software to increase its automation and accessibility and recommend that practitioners consider its use in population monitoring based on the identification of individuals in imagery.
README: Supplementing human observation with artificial intelligence impacts demographic estimates for a Critically Endangered lizard.
https://doi.org/10.5061/dryad.bvq83bkkh
Description of the data and file structure
These files contain two sets of encounter histories constructed by alternative individual identification methods, and were used in our paper to highlight the practical impact individual identification methods can have on real-world conservation scenarios. The collapsed encounter histories in these sets were used for goodness of fit testing whilst the full encounter histories formed the primary date for our analysis; using a robust design method in ProgramMARK.
Files and variables
The encounter history files are provided in .inp format; these can be opened and viewed in software capable of reading .txt files, such as notepad. In the 'collapsed' files, presence and absence of an individual is recorded for each of the 5 primary capture occasions, e.g.
/* 211 */ 11000 1;
where /* 211 */ indicates to ProgramMARK that 211 is the identifier for the individual the encounter history belongs to, 11000 indicates that this individual was captured within the first and second primary occasions, but not the following three, and 1; indicates that this is the capture pattern for one individual and that all information following will relate to a new individual.
Full encounter history files are formatted in the same way but record presences and absences across both primary and secondary occasions, running to a total of 52 occasions.
Further information regarding data formatting and analysis with ProgramMARK can be found here: Program MARK
File: AI_SupplementedEncounterHistory.inp
Description: Encounter history in .inp format suitable for use with ProgramMARK. Used for Robust Design Analysis of the AI-Supplemented data set.
File: AI_SupplementedEncounterHistory_Collapsed.inp
Description: Collapsed encounter history in .inp format suitable for use with ProgramMARK. Used for goodness of fit testing of the AI-Supplemented data set.
File: HumanOnlyEncounterHistory.inp
Description: Encounter history in .inp format suitable for use with ProgramMARK. Used for Robust Design Analysis of the Human-Only data set.
File: HumanOnlyEncounterHistory_Collapsed.inp
Description: Collapsed encounter history in .inp format suitable for use with ProgramMARK. Used for goodness of fit testing of the Human-Only data set.
Code/software
Data is formatted for analysis in ProgramMARK.
Methods
This data includes encounter histories saved as text .inp files for use in ProgramMARK. The collapsed encounter histories were used in goodness of fit tests (TESTS 2+3 in ProgramMARK). The full histories were used for Robust Design capture-recapture analysis (Huggin's p's and c's in ProgramMARK).