基于RBF模型的太湖北部叶绿素a浓度定量遥感反演

Quantitative retrieval of chlorophyll-a concentration in northern part of Lake Taihu based on RBF model

以大型内陆浅水湖泊一太湖为例,采用RBF(Radial Basis Function)神经网络建立该研究区域的叶绿素a浓度与同步影像数据的反演模型,较分析现今应用最广泛使用的BP(Back Propagation)神经网络模型,并通过对模型的验证、稳定性和鲁棒性分析评价了两种模型的泛化能力。结果表明,常规的BP神经网络模型收敛速度慢,极容易陷入局部最优解,而RBF比BP模型有更加优异的函数逼近、分类和模式识别能力,对反演叶绿素a浓度具有很强的泛化能力。

Taihu Lake,a large shallow inland lake,was taken as an example in this study. A quantitative inversion model was established between chlorophyll-a concentration and synchronous remote sensing images by using an RBF neural network. A comparison of this model with a BP neural network model,which has been extensively used in recent works,enabled a comparative analysis of the performance of BP and RBF in terms of validation,stability,and robustness. The results showed that the BF neural network has several shortcomings,such as slow convergence and easy fall into the local minima. In contrast,the RBF neural network has many advantages; for instance,it has stronger abilities of function approximation,classification,and pattern recognition than the BP network. The RBF model provides a good generalization from the perspective of inversing the chlorophylla concentration.