A user-friendly forest model with a multiplicative mathematical structure: a Bayesian approach to calibration
A user-friendly forest model with a multiplicative mathematical structure: a Bayesian approach to calibration
复制标题
DOI：
10.5194/gmdd-7-6997-2014
复制
发表时间：
2014-10
期刊：
影响因子：
--
通讯作者：
M. Bagnara;M. Oijen;D. Cameron;D. Gianelle;F. Magnani;M. Sottocornola
中科院分区：
文献类型：
--
作者：
M. Bagnara;M. Oijen;D. Cameron;D. Gianelle;F. Magnani;M. Sottocornola
森林模型正越来越多地用于研究生态系统的功能，通过复制世界各地非常不同的森林的碳通量和生产力。在过去的二十年中，需要简单和“易于使用”的模型，其特点是很少的参数和方程的实际应用，已经变得明显，并已开发了一些为此目的。这些模型旨在代表森林生态系统进程的主要驱动因素，同时适用于尽可能广泛的森林生态系统。最近，人们也越来越清楚，模型性能的评估不应只在准确性的估计和预测，而且在模型的不确定性估计。因此，贝叶斯方法已越来越多地应用于校准森林模型，其目的是估计其结果的不确定性，并比较其性能。一些被认为方便用户的森林模型依赖乘法或准乘法数学结构，这在校准过程中会造成问题，主要是因为参数之间的高度相关性。在使用马尔可夫链蒙特卡罗抽样的贝叶斯框架中，这可能会损害链的适当收敛和从正确的后验分布的抽样。在这里，我们展示了两种方法，以达到适当的收敛时，使用森林模型的乘法结构，应用不同的算法，不同的迭代次数在马尔可夫链蒙特卡罗或两步校准。结果表明，最近提出的算法自适应校准不赋予一个明显的优势，在这里使用的森林模型的大都市黑斯廷斯随机游走算法。此外，校准仍然耗时且数学上困难，因此由于获得可靠结果所需的校准过程，使用快速且用户友好的模型的优点可能会丧失。
Forest models are being increasingly used to study ecosystem functioning, through the reproduction of carbon fluxes and productivity in very different forests all over the world. Over the last two decades, the need for simple and “easy to use” models for practical applications, characterized by few parameters and equations, has become clear, and some have been developed for this purpose. These models aim to represent the main drivers underlying forest ecosystem processes while being applicable to the widest possible range of forest ecosystems. Recently, it has also become clear that model performance should not be assessed only in terms of accuracy of estimations and predictions, but also in terms of estimates of model uncertainties. Therefore, the Bayesian approach has increasingly been applied to calibrate forest models, with the aim of estimating the uncertainty of their results, and of comparing their performances. Some forest models, considered to be user-friendly, rely on a multiplicative or quasimultiplicative mathematical structure, which is known to cause problems during the calibration process, mainly due to high correlations between parameters. In a Bayesian framework using a Markov Chain Monte Carlo sampling this is likely to impair the reaching of a proper convergence of the chains and the sampling from the correct posterior distribution. Here we show two methods to reach proper convergence when using a forest model with a multiplicative structure, applying different algorithms with different number of iterations during the Markov Chain Monte Carlo or a two-steps calibration. The results showed that recently proposed algorithms for adaptive calibration do not confer a clear advantage over the Metropolis–Hastings Random Walk algorithm for the forest model used here. Moreover, the calibration remains time consuming and mathematically difficult, so advantages of using a fast and user-friendly model can be lost due to the calibration process that is needed to obtain reliable results.