模拟退火优化BP神经网络进行云相态分类

Simulated annealing optimized BP-ANN method for cloud thermodynamic phase retrieval

为利用中分辨率成像光谱仪(MODIS)多通道云图数据对云相态进行分类,提出了一种利用云图红外、可见光谱段数据,基于SABP神经网络的云相态分类算法。算法基于BP人工神经网络学习算法及模拟退火算法。在两算法结合的基础上,作出了有益于云相态分类的改进。基准网络模型为3层前馈BP网络,其中,隐层和输出层使用不同的激活函数。选取了5种光谱特征作为网络输入,输出两单元判定结果。为加速收敛速度,避免振荡,采用了一种自适应的变学习系数、惯性系数策略。个例分析表明,算法在中低纬度地区效果良好。

In order to retrieve cloud thermodynamic phase using MODIS images,an ANN-based method utilizing MODIS visible and infrared bands data and multi-spectral features was put forward.This method combines the relatively mature BP artificial neural network study algorithm and simu1ated annealing algorithm.Special amendments were made for better cloud thermodynamic phase retrieval performance. A triple-layer feedforward BP-ANN base model which used diverse activation functions was selected.Five spectral features were selected as input of the network model.the network output layer has two units for speeding up and avoiding vibration errors.A self-adapted variational study coefficient and inertia coefficient strategy was adopted.Case show presented the efficiency of the method in mid or low latitude regions.