Which modelling choices are supported when marine survey data are limited? An evidence map of 95 studies and a benchmark on structured subsamples of a trawl survey
Abstract
1. Marine management decisions increasingly rest on species distribution models fitted to sparse, clustered survey records, while the guidance for fitting them comes from data-rich settings. We asked which modelling choices have been tested under the conditions marine practitioners face, and whether the ordering that guidance assumes, spatial structure before algorithm choice, holds at the sample sizes at which it is applied. 2. We screened 330 records and retained 95 peer-reviewed marine applications that treated spatial autocorrelation explicitly, coding the contrasts tested, the sample sizes reported and the metric family. We then benchmarked, on the NWFSC West Coast Groundfish Bottom Trawl Survey, subsamples of three demersal species at 30 to 500 records crossed with dispersed and clustered coverage and with interpolation and extrapolation targets, fitting random forests and generalised additive models with and without spatial terms and a Gaussian Markov random field, evaluated against the held-out survey and by random and spatial block cross-validation. 3. No marine study crosses spatial treatment against algorithm family on the same data, and one reports performance below 100 records. In the benchmark, spatial structure helped or harmed according to how it was added: a coordinate smooth in a GAM reduced held-out AUC by 0.06 to 0.15 below 200 records and at every size under clustered sampling, a Matérn random field recovered most of that loss, and coordinates in a random forest helped only with dispersed coverage and an interpolation target. Dispersed coverage was worth as much as the gain from 30 to 500 records. Random cross-validation erred by up to 0.13 AUC in either direction according to coverage, and spatial blocking moved the estimate by less than 0.03. Likelihood models were overconfident at 30 to 50 records; a random forest with coordinates was nearly calibrated. 4. Spatial against non-spatial is not a contrast on which to base a recommendation; coverage is the first-order decision; internal validation does not reveal extrapolation loss when the sample is a band; and at the smallest sizes model choice affects the probability scale more than the ranking. The corpus and benchmark yield a practitioner guide graded by the evidence behind each element.
Related
- Modelling interpretable patient-level representationsfrom structured and simple multimodal data
- Anterior part of fronto-temporal language network as the neural basis of morphologically structured word form processing: evidence from Korean morphologically complex word.
- Metabolic collapse as a mechanism of developmental regression: convergent evidence from Kleefstra syndrome FDG-PET/CT imaging and Drosophila modelling
- Evidence-informed dynamic modelling and intervention prioritisation in diabetic bone marrow
- Intercepting hepatocellular carcinoma in chronic hepatitis B: a structured critical review of Chinese herbal medicine and the immune–fibrotic evidence
BibTeX
@misc{afrifayamoah2026which,
title = {{Which modelling choices are supported when marine survey data are limited? An evidence map of 95 studies and a benchmark on structured subsamples of a trawl survey}},
author = {Ebenezer Afrifa-Yamoah and Alexandre C Siquerira and Yaw Kwaafo Awuah-Mensah and Francky Fouedjio and Ute Mueller},
howpublished = {EcoEvoRxiv},
year = {2026},
month = sep,
doi = {10.32942/x2jm49},
url = {https://ecoevorxiv.org/repository/view/15074/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Which modelling choices are supported when marine survey data are limited? An evidence map of 95 studies and a benchmark on structured subsamples of a trawl survey
AU - Ebenezer Afrifa-Yamoah
AU - Alexandre C Siquerira
AU - Yaw Kwaafo Awuah-Mensah
AU - Francky Fouedjio
AU - Ute Mueller
PY - 2026
DA - 2026/09/23
PB - EcoEvoRxiv
DO - 10.32942/x2jm49
UR - https://ecoevorxiv.org/repository/view/15074/
AB - 1. Marine management decisions increasingly rest on species distribution models fitted to sparse, clustered survey records, while the guidance for fitting them comes from data-rich settings. We asked which modelling choices have been tested under the conditions marine practitioners face, and whether the ordering that guidance assumes, spatial structure before algorithm choice, holds at the sample sizes at which it is applied. 2. We screened 330 records and retained 95 peer-reviewed marine applications that treated spatial autocorrelation explicitly, coding the contrasts tested, the sample sizes reported and the metric family. We then benchmarked, on the NWFSC West Coast Groundfish Bottom Trawl Survey, subsamples of three demersal species at 30 to 500 records crossed with dispersed and clustered coverage and with interpolation and extrapolation targets, fitting random forests and generalised additive models with and without spatial terms and a Gaussian Markov random field, evaluated against the held-out survey and by random and spatial block cross-validation. 3. No marine study crosses spatial treatment against algorithm family on the same data, and one reports performance below 100 records. In the benchmark, spatial structure helped or harmed according to how it was added: a coordinate smooth in a GAM reduced held-out AUC by 0.06 to 0.15 below 200 records and at every size under clustered sampling, a Matérn random field recovered most of that loss, and coordinates in a random forest helped only with dispersed coverage and an interpolation target. Dispersed coverage was worth as much as the gain from 30 to 500 records. Random cross-validation erred by up to 0.13 AUC in either direction according to coverage, and spatial blocking moved the estimate by less than 0.03. Likelihood models were overconfident at 30 to 50 records; a random forest with coordinates was nearly calibrated. 4. Spatial against non-spatial is not a contrast on which to base a recommendation; coverage is the first-order decision; internal validation does not reveal extrapolation loss when the sample is a band; and at the smallest sizes model choice affects the probability scale more than the ranking. The corpus and benchmark yield a practitioner guide graded by the evidence behind each element.
ER -