Automated calibration strategies in ecological modeling using high-frequency in situ data
Résumé
Lake ecosystems are subject to various stressors (e.g. climate change, pollution, eutrophication) that threaten their ecological functioning, often leading to harmful algal blooms. Coupled hydrodynamic-ecological models are valuable tools to reach a deeper understanding of the key factors triggering events such as cyanobacterial blooms. However, their calibration and validation is often a challenging task. In situ data available for calibration are usually sparse in space and time and the high number of variables and interactions in the biogeochemical cycle is often reflected in models with a high number of parameters to be estimated.
High-frequency in situ data can help overcome these issues as they make it possible to perform the calibration on a shorter period for which the simulation time is not too long. Thus, we can use calibration methods that require performing an important number of model simulations for different parameter sets. Among these techniques, likelihood-free methods based on Bayesian inference are increasingly used in the fields of ecology and biology. Approximate Bayesian Computation (ABC) is a class of computational methods that can be applied to estimate most probable parameters without the need to evaluate the likelihood function. In our study, we apply and compare different calibration strategies based on ABC methods to calibrate and validate a 3D ecological model (Delft3D/BLOOM, Deltares). The case-study is a small urban lake (Greater Paris, France) which is affected by repeated cyanobacterial blooms and for which high-frequency in situ data are available.