Validation approaches of a geolocation framework to reconstruct movements of demersal fish equipped with data storage tags in a stratified environment.

• Validation of geolocation models should be cost-effective and reported systematically. • Stratification, heterogeneous bathymetry and vertical movements improve geolocation. • The geolocation framework resulted in an uncertainty of less than 20 km on average. • Incorrectly modelled start and end positions are strong criteria to discard tracks. Geolocation models are essential tools to derive detailed information on the movement patterns of demersal fish equipped with data storage tags recording temperature and depth. However, modelled horizontal positions are accompanied by an often unknown error which needs to be assessed to support the appropriate interpretation. We conducted five validation experiments in the southern Baltic Sea comprising a) simulated tracks, b) stationary, nearshore moored tags, c) temperature data from an automated offshore measuring station, d) a temperature-depth probe mounted to an otter board of a commercial vessel and e) tags mounted to the conductivity-temperature-depth probe (CTD) and the otter board of a survey vessel to obtain temperature-depth data from known locations. Using depth-temperature data from each experiment, we calculated modelled positions using an adapted R-package HMMoce model framework. These experiments allowed us to compare true and modelled positions to quantify model performance and assess its potential efficacy for future geolocation studies on tagged cod. Correctly modelled start and end positions were a good indicator for all validation methods; modelled and true tracks were generally similar with differences in daily positions varying on average between 2 and 19 km. The performance of the geolocation model was improved when temperature-depth profiles indicated heterogeneity in the oceanographic environment, particularly in stratified water columns. The validation methods confirmed that the modified HMMoce framework could be used as a geolocation tool for Eastern Baltic cod when an uncertainty of less than 20 km from the true position is acceptable. Since all presented methods are cost-effective, transparent, open source and reproducible, they can be routinely applied to understand expected model uncertainties and thus make more robust inference about movements of tagged fish. [ABSTRACT FROM AUTHOR]