Is my model fit for purpose? Validating a population model for predicting freshwater fish responses to flow management

Models based on ecological processes ("process-explicit models") are often used to predict ecosystem responses to environmental changes or management scenarios. However, models are imperfect and need to be validated, ideally by testing their assumptions and outputs against independent empirical data sets. Examples of validation of process-explicit models are rare. Recently, stochastic population models have been developed to predict the likely responses (over 10-120 years) of a riverine fish (golden perch, Macquaria ambigua) to flow management in the Murray-Darling Basin (MDB) in eastern Australia, one of the world's most regulated river basins. Declines of golden perch (and other species) are a direct consequence of altered hydrology, and managers require information to predict how fish will respond to possible future hydrological conditions to guide the substantial investments in flow management. Here, we use two independent field data sets to validate our population model. We compared model predictions to observed trends to ask: (1) how do predicted population sizes and growth rates compare to observed data? (2) does the correlation between predicted and observed population sizes and growth rates vary among populations? (3) does the correlation between predicted and observed population sizes and growth rates vary across observed hydrological conditions? and (4) how do modelled and observed fish movement rates compare? We found reasonable correlations between fish population sizes and growth rates as predicted by the model and observed in independent data sets for several populations (Aim 1) but the strength of these correlations varied among populations (Aim 2) and hydrological conditions (Aim 3). Predicted and observed fish movement rates were strongly correlated (Aim 4). Population models are frequently used in conservation decision-making but are rarely validated. We demonstrate that: (1) validation can identify model strengths and weaknesses; (2) observed data sets often have inherent limitations that can preclude robust validations; (3) validation is likely be more common if appropriate observed data sets are available; and (4) validation should consider the purpose of modelling. Wider consideration of these messages would contribute to more critical examinations of models so they can be most appropriately used in conservation decision-making.

基于生态过程的模型（“显式过程模型”，process-explicit models）常被用于预测生态系统对环境变化或管理情景的响应。然而模型并非完美无缺，需要开展验证工作，理想方式是通过独立经验数据集检验其假设与输出结果。目前显式过程模型的验证案例较为罕见。

近期，研究人员开发了随机种群模型（stochastic population models），用于预测澳大利亚东部墨累-达令盆地（Murray-Darling Basin, MDB）——全球受调控程度最高的流域之一——内的河栖鱼类黄金鲈（Macquaria ambigua）在未来10至120年间对水流管理的潜在响应。黄金鲈（及其他物种）的种群衰退是水文情势改变的直接后果，管理者需要相关信息以预测鱼类对未来可能的水文条件的响应，从而指导水流管理领域的大额投资。

本研究使用两套独立的野外数据集对本研究的种群模型进行验证。我们将模型预测结果与观测趋势进行对比，旨在回答以下四个问题：（1）预测的种群规模与生长速率与观测数据相比表现如何？（2）预测与观测的种群规模、生长速率之间的相关性是否因种群而异？（3）预测与观测的种群规模、生长速率之间的相关性是否随观测到的水文条件变化而改变？（4）模型模拟的鱼类移动速率与观测值相比表现如何？

研究结果显示，针对多个种群，模型预测的鱼类种群规模与生长速率与独立数据集的观测结果之间存在合理的相关性（研究目标1），但此类相关性的强度因种群（研究目标2）和水文条件（研究目标3）而异。模型预测与观测的鱼类移动速率之间存在显著相关性（研究目标4）。

种群模型常被应用于保护决策，但相关验证工作鲜少开展。本研究表明：（1）模型验证可识别其优势与不足；（2）观测数据集往往存在固有局限性，可能无法支撑稳健的验证工作；（3）若能获取合适的观测数据集，模型验证工作或将更为常见；（4）验证过程应考量建模的初衷。对这些要点的更广泛审视，将有助于对模型开展更严谨的评估，从而使其能更恰当地应用于保护决策中。