多阈值和神经网络卫星云图云系自动分割试验

AUTOMATIC SEGMENTATION OF SATELLITE IMAGE USING HIERARCHICAL THRESHOLD AND NEURAL NETWORK

卫星云图自动分割是实现卫星云图云系自动识别的基础 .选用 1 992～ 1 994年和 1 997～1 998年夏季有典型天气系统的 1 77幅 GMS红外云图建立了云系模型库 ,云系分类样本 30 79个 ,包含 1 6类云系 ,云系分割样本 2 764个 .利用云系分割样本集进行神经网络试验 ,训练集为从 32幅云图中抽取的 484个样本 ,测试集为从 1 45幅云图中抽取的 2 2 80个样本 ,神经网络模型训练正确率达到 98.8% ,测试正确率为 86.4% .用 1 997年 7月 1 8～ 2 1日和 1 998年 6月 1 5～ 1 7日的两组卫星云图做自动分割应用试验 ,结果经专家判识 ,正确率达到 90 %以上 .本文的工作表明 :用多阈值和人工神经网络相结合方法对卫星云图进行云分割在实际应用中是可行的 .卫星云图自动分割系统的输入是 GMS红外云图 ,输出是分割出的每一个云区 ,同时还包括云区的边界链码、起始点、周长、面积 ,并保留了原始图像数据 .在下一步的云系识别过程中 ,可以在此基础上进行云系分类识别试验 .

Automatic segmentation of satellite image is the base of the automatic identification of cloud systems. This paper presents a combined method of hierarchical threshold segmentation and neural network to segment the image into separate synoptic systems for identification purpose. The processes include selecting all potential TBB thresholds of cloud segments and plotting temperature contours on satellite images for each threshold, then selecting cloud regions which is identified by a certain temperature contour and best representation of synoptic systems, and finally forming a segmented image by combining these regions. How to select these regions by computer is an experiential and uncertain problem. In this paper it is solved by the up bottom and bottom up heuristic and neural network method.A cloud pattern database is established. It includes 177 GMS satellite infrared images with 16 kinds of typical synoptic systems in the summer of 1992-1994,1997-1998.There are 484 training samples from 32 satellite images and 2280 testing samples from 145 satellite images. The neural network accuracy rate for these training samples is 98.8% and 86.4% for the testing samples. The experiment accuracy rate for the application test is above 90% using testing images of 18-21 July 1997 and 15-17 June 1998. The input of the satellite image automatic segmentation system is GMS images, and the outputs include the boundary chain code, start location, perimeter and area of each region. These outputs are useful to further classification of cloud patterns.