关于卫星气象云图准确检测的仿真研究

Simulation Study on Accurate Detection of Satellite Meteorological Cloud Image

利用卫星云图进行气象应用研究,首先要对卫星云图解译,而云图检测分类是气象卫星云图解译的核心。对气象云图准确检测的准确率高低直接影响到后续的大气科学研究和气象预测应用,而传统的浅层学习分类算法用于卫星气象云图检测分类时,不能很好的对卫星光学参数和卫星云图进行特征表示,容易造成处理规模过大、分析过程复杂以及陷入局部极小值等问题,导致了云图检测不准确,云类别的误检率很高。而深度神经网络对学习样本数量要求较高,并且检测速度慢,在分类速度和分类精度上无法满足气象研究和应用需求。针对这些问题,利用优化过的多粒度级联森林对卫星云图进行检测,能充分的对云图进行特征表示,并且具有很好的泛化性能。在对中国HJ-1A/B卫星云图的实验结果表明,采用基于多粒度级联森林方法对卫星云图进行检测,可以很好的提取云图特征,并且能够进行较好的小样本学习。而且多粒度级联森林方法检测速度快,云图分类时厚云和薄云之间的过渡区域清晰,而且相比传统阈值法、卷积神经网络模型及深度极限学习机模型的云图识别准确率更高。

To study the meteorological applications of satellite cloud images,the first step is to interpret the satellite cloud images,and cloud map classification is the meteorological satellite cloud interpretation of the core. The accuracy of the accurate detection of meteorological cloud image directly affects the follow-up atmospheric science research and meteorological forecasting application,and the traditional shallow learning classification algorithm for satellite weather cloud image classification,can not be very good for satellite optical parameters and satellite cloud. The feature representation is easy to cause the processing scale to be too large. The analysis process is complicated and the local minimum value is lost,which leads to the inaccurate detection of the cloud and high false detection rate of the cloud category. The deep neural network requires large number of learning samples,and the detection speed is slow,so the classification speed and classification accuracy can not meet the meteorological research and application requirements. In view of these problems,the optimized multi-granularity cascade forest was used to detect the satellite cloud image,which can express the features of the cloud map and has good generalization performance. The experimental results of HJ-1A/B satellite cloud show that using the multi-granularity cascade forest to detect the satellite cloud image can extract the features of the cloud map well and carry out better small sample learning. The multi-granularity cascade forest method is fast and the transition area between cloud and thin clouds is clear,and the accuracy of cloud image recognition is higher than that of traditional threshold method,convolution neural network model and deep limit learning machine model.