arxiv: 2605.08532 · v1 · submitted 2026-05-08 · 📊 stat.AP · stat.ME

Recognition: no theorem link

Accounting for variable detection functions in temporal abundance modeling via transfer learning

Christopher K. Wikle, Erin M. Schliep, Kevin M. Collins, Tyler Wagner

Pith reviewed 2026-05-12 00:58 UTC · model grok-4.3

classification 📊 stat.AP stat.ME

keywords transfer learningcapture-recapturecatch-per-unit-effortabundance modelingdetection probabilitytemporal trendsfisheriessmallmouth bass

0 comments

The pith

Transfer learning from capture-recapture data allows catch-per-unit-effort models to account for variable detection probabilities, leading to improved abundance estimates and better detection of temporal trends.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Capture-recapture surveys provide detailed information on how detection probability varies with environmental conditions, but they are costly to conduct at large scales. Catch-per-unit-effort data are cheaper and more widely available but typically assume constant detection, which can bias abundance estimates. The paper introduces a transfer learning method that extracts the relationship between covariates and detection from CR data and applies it to adjust CPUE models. Simulations demonstrate that this yields more accurate abundance estimates and greater power to identify real population changes over time. The approach is illustrated with data from smallmouth bass populations in Pennsylvania rivers.

Core claim

We propose an approach to (i) learn the effect of environmental covariates on detection probabilities from CR data and (ii) transfer these detection functions to CPUE models for improved inference. Shown empirically through a simulation study, this approach improves estimates of abundance and the ability to detect temporal trends. We apply our transfer learning method using CR and CPUE data to recreationally important smallmouth bass fisheries in Pennsylvania, USA rivers.

What carries the argument

Transfer learning of detection functions estimated from capture-recapture data to catch-per-unit-effort abundance models, using shared environmental covariates.

Load-bearing premise

The relationship between environmental covariates and detection probability is sufficiently consistent between capture-recapture and catch-per-unit-effort sampling methods to allow direct transfer of the learned functions.

What would settle it

A validation study where abundance estimates from the transfer method are compared against known true abundances from a fully observed population or independent high-effort surveys, showing no improvement or worse performance than standard CPUE models.

Figures

Figures reproduced from arXiv: 2605.08532 by Christopher K. Wikle, Erin M. Schliep, Kevin M. Collins, Tyler Wagner.

**Figure 1.** Figure 1: Posteriors for naive (N˜ tj ; black) and transfer learning (N˜ ∗ tj ; red) abundance estimates over time for a single simulation run. Estimates are provided for two size classes of fish (size class 1 = < 200mm and size class 2 = > 200 mm total length). The true values of Ntj are marked with blue crosses. Note that detection probability is affected by environmental covariates for size class 1, but not for s… view at source ↗

**Figure 2.** Figure 2: Transfer learning model output for Juniata River segment 3 and Susquehanna [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗

**Figure 3.** Figure 3: Posterior distributions of daily detection probabilities ( [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗

**Figure 4.** Figure 4: Posterior distributions of annual abundance ( [PITH_FULL_IMAGE:figures/full_fig_p029_4.png] view at source ↗

**Figure 5.** Figure 5: Model output for Juniata River Segment 3 CPUE data. The top two panels are [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗

**Figure 6.** Figure 6: Model output for Juniata River Segment 4 CPUE data. The top two panels are [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗

**Figure 7.** Figure 7: Model output for Juniata River Segment 5 CPUE data. The top two panels are [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗

**Figure 8.** Figure 8: Model output for Juniata River Segment 6 CPUE data. The top two panels are [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗

**Figure 9.** Figure 9: Model output for Juniata River Segment 7 CPUE data. The top two panels are [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗

**Figure 10.** Figure 10: Model output for North Branch Susquehanna River Segment 1 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p035_10.png] view at source ↗

**Figure 11.** Figure 11: Model output for North Branch Susquehanna River Segment 3 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p036_11.png] view at source ↗

**Figure 12.** Figure 12: Model output for North Branch Susquehanna River Segment 5 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p037_12.png] view at source ↗

**Figure 13.** Figure 13: Model output for North Branch Susquehanna River Segment 6 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p038_13.png] view at source ↗

**Figure 14.** Figure 14: Model output for North Branch Susquehanna River Segment 8 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p039_14.png] view at source ↗

**Figure 15.** Figure 15: Model output for North Branch Susquehanna River Segment 9 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p040_15.png] view at source ↗

**Figure 16.** Figure 16: Model output for North Branch Susquehanna River Segment 10 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p041_16.png] view at source ↗

**Figure 17.** Figure 17: Model output for Susquehanna River Segment 2 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p042_17.png] view at source ↗

**Figure 18.** Figure 18: Model output for Susquehanna River Segment 3 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p043_18.png] view at source ↗

**Figure 19.** Figure 19: Model output for Susquehanna River Segment 4 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p044_19.png] view at source ↗

**Figure 20.** Figure 20: Model output for Susquehanna River Segment 5 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p045_20.png] view at source ↗

**Figure 21.** Figure 21: Model output for Susquehanna River Segment 6 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p046_21.png] view at source ↗

**Figure 22.** Figure 22: Model output for Susquehanna River Segment 7 CPUE data. The top two [PITH_FULL_IMAGE:figures/full_fig_p047_22.png] view at source ↗

**Figure 23.** Figure 23: Model output for West Branch Susquehanna River Segment 10 CPUE data. [PITH_FULL_IMAGE:figures/full_fig_p048_23.png] view at source ↗

read the original abstract

Relative abundance, measured as the number of animals caught per unit of sampling effort (CPUE), is commonly used to monitor fish and wildlife populations, largely because sampling methods are cost-effective to implement. Modeling relative abundance, however, requires the assumption that the detection probability is constant across sampling events. This assumption is likely not valid, as the probability of detection often varies as a function of several factors, including the characteristics of individual animals and environmental conditions at the time of sampling. In contrast, methods to estimate absolute abundance, such as capture-recapture (CR), account for variable detection, but are often infeasible to implement across large spatiotemporal scales. Despite this, CR data are sometimes available for species of interest, albeit at smaller spatiotemporal extents. Leveraging information on detection probabilities from CR data to help inform estimates of widely available CPUE data could strengthen inferences about the status of fish and wildlife populations. We propose an approach to (i) learn the effect of environmental covariates on detection probabilities from CR data and (ii) transfer these detection functions to CPUE models for improved inference. Shown empirically through a simulation study, this approach improves estimates of abundance and the ability to detect temporal trends. We apply our transfer learning method using CR and CPUE data to recreationally important smallmouth bass (\textit{Micropterus dolomieu}) fisheries in Pennsylvania, USA rivers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper proposes a transfer learning pipeline to borrow detection functions from CR data for CPUE abundance models, but the simulations skip testing whether covariate effects actually match across the two sampling contexts.

read the letter

The main takeaway is that this work tries to fix a known bias in CPUE data by learning how covariates affect detection from limited capture-recapture samples and plugging those functions into temporal abundance models. The abstract presents this as a new approach for improving trend detection in relative abundance time series. The simulation shows gains in abundance estimates and trend power, and they walk through an application to smallmouth bass in Pennsylvania rivers using both data types. That practical focus on leveraging scarce CR data for common monitoring is the part that lands. The method is clearly motivated by real constraints in fisheries and wildlife work. The soft spot is the simulation design. It generates both CR and CPUE from the identical detection process, so it never checks what happens when environmental effects on detection differ between the two protocols, which is the realistic case with different gears or spatial scales. The Pennsylvania example does not appear to test or adjust for that mismatch either. Without numbers on the size of the improvement or details on how the transfer is implemented, the central claim stays hard to weigh. This paper is aimed at statistical ecologists and managers who already collect some CR data alongside routine CPUE sampling and want a way to reduce detection bias in trend estimates. A reader working on fisheries monitoring or population trend analysis would see direct relevance. It deserves a serious referee because the problem is common and the proposal is usable, even if the transfer assumption needs more stress testing. I would send it out for review.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a transfer learning approach to estimate variable detection probabilities from limited capture-recapture (CR) data and transfer the resulting detection functions to inform catch-per-unit-effort (CPUE) models for absolute abundance estimation and temporal trend detection. The method is evaluated in a simulation study claimed to show improved abundance estimates and trend detection power, and is applied to smallmouth bass (Micropterus dolomieu) fisheries using CR and CPUE data from Pennsylvania rivers.

Significance. If the transfer assumption holds, the approach could substantially increase the value of abundant but detection-biased CPUE data by borrowing strength from more rigorous but spatially limited CR studies, improving the reliability of population monitoring for management. The simulation provides a proof-of-concept under matched conditions, but the real-data application would gain from explicit validation of transferred effects.

major comments (2)

[Simulation study] Simulation study section: the data-generating process uses the same detection function (identical covariate effects on p) for both CR and CPUE observations, so the reported improvements in abundance accuracy and trend detection do not address bias arising from protocol mismatch (gear type, effort distribution, spatial scale). This is load-bearing for the central claim that transferred functions improve real-world CPUE inference.
[Application to Pennsylvania smallmouth bass data] Application section: no diagnostic or validation is described to confirm that CR-derived covariate effects (e.g., on flow or temperature) match those operating in the CPUE electrofishing/angling samples; without this, the Pennsylvania results remain an untested extrapolation.

minor comments (2)

[Abstract] Abstract and results lack any numerical summary (bias reduction, MSE, power gain, or confidence intervals) of the claimed simulation improvements, hindering assessment of practical significance.
[Methods] Notation for the transferred detection function p(detection | covariates) and its plug-in into the CPUE likelihood should be defined explicitly with an equation in the methods.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and insightful comments on our manuscript. We address each major comment below and describe the revisions we will incorporate to strengthen the work.

read point-by-point responses

Referee: Simulation study section: the data-generating process uses the same detection function (identical covariate effects on p) for both CR and CPUE observations, so the reported improvements in abundance accuracy and trend detection do not address bias arising from protocol mismatch (gear type, effort distribution, spatial scale). This is load-bearing for the central claim that transferred functions improve real-world CPUE inference.

Authors: We agree that the simulation evaluates performance under the assumption of transferable detection functions and does not directly quantify bias from protocol mismatch. This design isolates the benefit of incorporating variable detection when transfer is valid, which is the method's core premise. To address the concern, we will add a new simulation scenario with mismatched covariate effects on detection (e.g., differing slopes for flow and temperature, plus protocol-specific noise) between CR and CPUE data-generating processes. Results will show performance degradation under mismatch and conditions where transfer remains advantageous. We will also revise the methods and discussion to emphasize that real-world gains require the transfer assumption and to recommend diagnostics for mismatch. revision: yes
Referee: Application section: no diagnostic or validation is described to confirm that CR-derived covariate effects (e.g., on flow or temperature) match those operating in the CPUE electrofishing/angling samples; without this, the Pennsylvania results remain an untested extrapolation.

Authors: We acknowledge that the current application lacks explicit validation of transferred effects. In the revision we will add a diagnostic subsection that compares the posterior means and credible intervals for shared covariate effects (flow, temperature) estimated from the CR model against those recovered in the CPUE model after transfer. We will also articulate the biological and methodological rationale for transferability, noting that both datasets were collected in comparable Pennsylvania river habitats using gears sensitive to the same environmental drivers. The discussion will explicitly state the transfer assumption, its unverifiable aspects with available data, and the need for future cross-validation studies. revision: yes

Circularity Check

0 steps flagged

No circularity in transfer learning from independent CR data to CPUE models

full rationale

The paper derives its method by estimating covariate effects on detection probability from separate capture-recapture (CR) datasets and transferring those functions as plug-ins into CPUE likelihoods. This uses distinct data sources rather than fitting parameters on a subset and relabeling a related quantity from the same data as a prediction. The simulation generates both data types from a shared ground-truth model solely to benchmark recovery performance, not to create a self-referential result. No self-citations, uniqueness theorems, or ansatz smuggling appear in the derivation chain. The claims rest on external empirical validation and application to Pennsylvania river data, rendering the approach self-contained against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that detection covariate effects are transferable between CR and CPUE sampling; no free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption Detection probability varies as a function of environmental covariates and these functions can be learned from CR data and transferred to CPUE models.
This is the core premise enabling the transfer learning step described in the abstract.

pith-pipeline@v0.9.0 · 5552 in / 1244 out tokens · 42357 ms · 2026-05-12T00:58:37.297375+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

[1]

The Condor , volume=

Factors affecting detection probability, effective area surveyed, and species misidentification in grassland bird point counts , author=. The Condor , volume=. 2019 , publisher=

work page 2019
[2]

Journal of Caribbean Ornithology , volume=

Effects of observer skill and survey method on forest bird abundance data: Recommendations for citizen science conservation monitoring in the Caribbean , author=. Journal of Caribbean Ornithology , volume=

work page
[3]

Canadian Journal of Fisheries and Aquatic Sciences , volume=

Predicting fish species richness and habitat relationships using Bayesian hierarchical multispecies occupancy models , author=. Canadian Journal of Fisheries and Aquatic Sciences , volume=. 2020 , publisher=

work page 2020
[4]

Transactions of the American Fisheries Society , volume=

Evaluation of electrofishing catch per unit effort for indexing fish abundance in Florida lakes , author=. Transactions of the American Fisheries Society , volume=. 2013 , publisher=

work page 2013
[5]

Fisheries Management and Ecology , volume=

Accounting for variation in species detection in fish community monitoring , author=. Fisheries Management and Ecology , volume=. 2014 , publisher=

work page 2014
[6]

North American Journal of Fisheries Management , volume=

Validation of abundance estimates from mark--recapture and removal techniques for rainbow trout captured by electrofishing in small streams , author=. North American Journal of Fisheries Management , volume=. 2005 , publisher=

work page 2005
[7]

Freshwater Biology , volume=

Imperfect detection and the determination of environmental flows for fish: challenges, implications and solutions , author=. Freshwater Biology , volume=. 2016 , publisher=

work page 2016
[8]

Fisheries , volume=

Toward improved understanding of streamflow effects on freshwater fishes , author=. Fisheries , volume=. 2022 , publisher=

work page 2022
[9]

Transactions of the American Fisheries Society , volume=

Factors affecting nesting success of smallmouth bass in a regulated Virginia stream , author=. Transactions of the American Fisheries Society , volume=. 1995 , publisher=

work page 1995
[10]

Canadian Journal of Fisheries and Aquatic Sciences , volume=

How reliable are the abundance indices derived from commercial catch--effort standardization? , author=. Canadian Journal of Fisheries and Aquatic Sciences , volume=

work page
[11]

Fisheries Research , volume=

Indices of abundance in the Gulf of Mexico reef fish complex: A comparative approach using spatial data from vessel monitoring systems , author=. Fisheries Research , volume=. 2018 , publisher=

work page 2018
[12]

Biometrics , volume=

Hierarchical spatiotemporal matrix models for characterizing invasions , author=. Biometrics , volume=. 2007 , publisher=

work page 2007
[13]

Fisheries , volume=

Trends in inland commercial fisheries in the United States , author=. Fisheries , volume=. 2020 , publisher=

work page 2020
[14]

Fish and Fisheries , volume=

A fresh look at inland fisheries and their role in food security and livelihoods , author=. Fish and Fisheries , volume=. 2019 , publisher=

work page 2019
[15]

Biometrika , volume=

A note on the multiple-recapture census , author=. Biometrika , volume=. 1965 , publisher=

work page 1965
[16]

Biometrics , pages=

Mark-recapture models with parameters constant in time , author=. Biometrics , pages=. 1982 , publisher=

work page 1982
[17]

Fisheries Research , volume=

Accounting for uncertainty when estimating drivers of imperfect detection: An integrated approach illustrated with snorkel surveys for riverine fishes , author=. Fisheries Research , volume=. 2022 , publisher=

work page 2022
[18]

PloS one , volume=

Accounting for imperfect detection in ecology: a quantitative review , author=. PloS one , volume=. 2014 , publisher=

work page 2014
[19]

Environmental Science and Pollution Research , volume=

Evaluating and validating abundance monitoring methods in the absence of populations of known size: review and application to a passive tracking index , author=. Environmental Science and Pollution Research , volume=. 2015 , publisher=

work page 2015
[20]

Environmetrics: The official journal of the International Environmetrics Society , volume=

Large scale wildlife monitoring studies: statistical methods for design and analysis , author=. Environmetrics: The official journal of the International Environmetrics Society , volume=. 2002 , publisher=

work page 2002
[21]

North American Journal of Fisheries Management , volume=

Adapting standardized trout monitoring to a changing climate for the upper Yellowstone River, Montana, USA , author=. North American Journal of Fisheries Management , volume=. 2024 , publisher=

work page 2024
[22]

Fisheries , volume=

A fishery after the decline: the Susquehanna River Smallmouth Bass story , author=. Fisheries , volume=. 2020 , publisher=

work page 2020
[23]

Analysis and interpretation of freshwater fisheries data

Relative abundance and catch per unit effort , author=. Analysis and interpretation of freshwater fisheries data. American Fisheries Society, Bethesda, Maryland , pages=

work page
[24]

Biometrika , volume=

A Bayesian analysis of multiple-recapture sampling for a closed population , author=. Biometrika , volume=. 1981 , publisher=

work page 1981
[25]

Ecology , volume=

An integrated data model to estimate abundance from counts with temporal dependence and imperfect detection , author=. Ecology , volume=. 2025 , publisher=

work page 2025
[26]

Journal of Applied Ecology , volume=

Accounting for spatiotemporal sampling variation in joint species distribution models , author=. Journal of Applied Ecology , volume=. 2024 , publisher=

work page 2024
[27]

Journal of Big data , volume=

A survey of transfer learning , author=. Journal of Big data , volume=. 2016 , publisher=

work page 2016
[28]

arXiv preprint arXiv:2312.13484 , year=

Bayesian transfer learning , author=. arXiv preprint arXiv:2312.13484 , year=

work page arXiv
[29]

Journal of Statistical Software , volume=

Estimating population abundance using sightability models: R SightabilityModel package , author=. Journal of Statistical Software , volume=

work page
[30]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Characterizing and avoiding negative transfer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[31]

2020 , publisher=

Joint species distribution modelling: With applications in R , author=. 2020 , publisher=

work page 2020
[32]

, author=

Rank correlation methods. , author=. 1948 , publisher=

work page 1948