Ground-based Visible Cloud Image Classification Method Based on KNN Algorithm
-
摘要: 云图的自动分类是实现地基云自动化观测的技术保障。该文探讨了一种先将云图分为积状云、层状云和卷云3大类的分类方案,通过对3大云类和晴空这4种天空类型的纹理特征、颜色特征和形状特征进行分析,选取了21个特征参量,并采用K最近邻分类器 (K-Nearest Neighbor,KNN), 在不同的K取值情况下对这几类天空类型进行了分类识别。结果表明:新的分类方案是可行的,且当纹理特征、颜色特征和形状特征结合使用时获取了比单独利用纹理特征、颜色特征和形状特征以及它们两两组合时更好的识别效果。当K=7且使用21个特征参量时,KNN算法对积状云、层状云、卷云和晴空的识别最好, 识别正确率分别为91.1%,74.4%,70.0%和100.0%,平均正确率为83.9%。Abstract: Cloud plays an important role in the meteorological research, and it is one of the most important factors of earth's energy balance and hydrological cycle. In order to actualize the automatic ground-based observation of clouds, automatic classification of cloud image is a difficult problem.A cloud classification scheme which classifies the cloud images into cumulus, stratus and cirrus is discussed. The clear sky is considered as a separate category in the scheme. Three kinds of image features, texture, color and shape are analyzed. The texture features describe the local information of image by using gray information normally, which have the characteristics for translation invariance. The color features consider the color of the image and focus on description of the overall image information, which have the characteristics for translation, rotation and scale invariability. The shape features describe the outline or region feature of the specific objectives and focus on description of single target. By analyzing the cloud image features of four different sky conditions, extraction algorithms are introduced in details. Using gray-level co-occurrence matrix and Tamura texture, color moment, and moment invariants, 21 characteristic parameters are extracted. Because of its high performance in solving complex issues, simplicity of implementation and low computational complexity, the K-Nearest Neighbor (KNN) classification algorithm is selected to process 21 characteristic parameters. 8 different K values and different features combination are used to recognize the 4 types of sky conditions. Classification experiments are conducted using single feature, combination of each two features, and all of these features together. The 7 experimental results demonstrate that the new scheme is feasible. And using texture features, color features and shape features together can get better performance than using these features alone or any two of them combined. When the parameter K is set to 7 and all 21 characteristic parameters are considered, the identification accuracy of cumulus, stratus, cirrus and clear sky are 91.1%, 74.4%, 70.0% and 100.0%, respectively, with the average accuracy up to 83.9%.
-
Key words:
- texture features;
- color features;
- shape features;
- KNN;
- cloud classification
-
表 1 云图分类表
Table 1 Types of cloud image
分类 积状云 层状云 卷云 高云 卷积云 卷层云 卷云 中云 高积云
(蔽光高积云除外)高层云,
蔽光高积云低云 积云,积雨云,
积云性层积云,堡状层
积云,荚状层积云层云,雨层云,
透光层积云,
蔽光层积云表 2 云图特征量平均值统计表
Table 2 Average values of characteristic parameters
特征量 积状云 层状云 卷云 晴空 Tamura粗糙度 0.8911 0.8697 0.8652 0.6877 Tamura对比度 0.3967 0.2079 0.2737 0.0941 Tamura方向度 0.4167 0.3057 0.3776 0.0398 GLCM对比度 0.1865 0.1535 0.1614 0.1471 GLCM相关 0.9774 0.9282 0.9617 0.7906 GLCM能量 0.3464 0.5426 0.4382 0.7296 GLCM局部平稳 0.9841 0.9869 0.9862 0.9930 GLCM熵 0.4183 0.5546 0.3698 0.2945 色调 (一阶分量) 0.5987 0.3735 0.6377 0.6628 饱和度 (一阶分量) 0.3237 0.1717 0.4548 0.6817 亮度 (一阶分量) 0.6561 0.7358 0.6617 0.6565 色调 (二阶分量) 0.2154 0.2841 0.0985 0.0070 饱和度 (二阶分量) 0.4041 0.1201 0.3183 0.0996 亮度 (二阶分量) 0.4429 0.2905 0.2790 0.1488 不变矩1 0.6811 0.6798 0.6990 0.7125 不变矩2 0.2807 0.3230 0.3148 0.3017 不变矩3 0.3445 0.2352 0.3579 0.2435 不变矩4 0.2943 0.2036 0.3063 0.2067 不变矩5 0.3256 0.2177 0.3106 0.2135 不变矩6 0.0097 0.0153 0.0096 0.0140 不变矩7 0.2382 0.1528 0.2355 0.1740 表 3 K=5时,提取纹理特征量的分类混淆矩阵
Table 3 Confusion matrix using texture features alone when K is set to 5
天空类型 积状云 层状云 卷云 晴空 积状云 57.8%(52) 21.1%(19) 21.1%(19) 0.0%(0) 层状云 7.8%(7) 72.2%(65) 13.3%(12) 6.7%(6) 卷云 25.6%(23) 22.2%(20) 51.1%(46) 1.1%(1) 晴空 1.1%(1) 10.0%(9) 1.1%(1) 87.8%(79) 注:括号内为相应的样本数。 表 4 K=11时,提取颜色特征量的分类混淆矩阵
Table 4 Confusion matrix using color features alone when K is set to 11
天空类型 积状云 层状云 卷云 晴空 积状云 82.2%(74) 5.6%(5) 12.2%(11) 0.0%(0) 层状云 11.1%(10) 77.8%(70) 10.0%(9) 1.1%(1) 卷云 24.4%(22) 8.9%(8) 58.9%(53) 7.8%(7) 晴空 0.0%(0) 0.0%(0) 1.1%(1) 98.9%(89) 注:括号内为相应的样本数。 表 5 K=1时,提取形状特征量的分类混淆矩阵
Table 5 Confusion matrix using shape features alone when K is set to 1
天空类型 积状云 层状云 卷云 晴空 积状云 32.2%(29) 28.9%(26) 31.1%(28) 7.8%(7) 层状云 20.0%(18) 41.1%(37) 18.9%(17) 20.0%(18) 卷云 26.7%(24) 21.1%(19) 35.6%(32) 16.7%(15) 晴空 3.3%(3) 15.6%(14) 13.3%(12) 67.8%(61) 注:括号内为相应的样本数。 表 6 K=51时,提取纹理和颜色特征量的分类混淆矩阵
Table 6 Confusion matrix when the texture features are used in conjunction with color features and K is set to 51
天空类型 积状云 层状云 卷云 晴空 积状云 86.7%(78) 6.7%(6) 6.7%(6) 0.0%(0) 层状云 11.1%(10) 76.7%(69) 11.1%(10) 1.1%(1) 卷云 23.3%(21) 6.7%(6) 68.9%(62) 1.1%(1) 晴空 0.0%(0) 0.0%(0) 1.0%(1) 98.9%(89) 注:括号内为相应的样本数。 表 7 K=31时,提取纹理和形状特征量的分类混淆矩阵
Table 7 Confusion matrix when the texture features are used in conjunction with shape features and K is set to 31
天空类型 积状云 层状云 卷云 晴空 积状云 61.1%(55) 21.1%(19) 17.8%(16) 0.0%(0) 层状云 8.9%(8) 65.6%(59) 17.8%(16) 7.8%(7) 卷云 22.2%(20) 18.9%(17) 55.6%(50) 3.3%(3) 晴空 0.0%(0) 7.8%(7) 2.2%(2) 90.0%(81) 注:括号内为相应的样本数。 表 8 K=31时,提取颜色和形状特征量的分类混淆矩阵
Table 8 Confusion matrix when the color features are used in conjunction with shape features and K is set to 31
天空类型 积状云 层状云 卷云 晴空 积状云 88.9%(80) 4.4%(4) 6.7%(6) 0.0%(0) 层状云 12.2%(11) 76.7%(69) 10.0%(9) 1.1%(1) 卷云 16.7%(15) 11.1%(10) 63.3%(57) 8.9%(7) 晴空 0.0%(0) 0.0%(0) 6.6%(6) 93.4%(84) 注:括号内为相应的样本数。 表 9 K=7时,提取纹、颜色和形状特征量的分类混淆矩阵
Table 9 Confusion matrix when the texture features are used in conjunction with color and shape features and K is set to 7
天空类型 积状云 层状云 卷云 晴空 积状云 91.1%(82) 5.6%(5) 3.3%(3) 0.0%(0) 层状云 12.2%(11) 74.4%(67) 11.1%(10) 2.2%(2) 卷云 20.0%(18) 7.8%(7) 70.0%(63) 2.2%(2) 晴空 0.0%(0) 0.0%(0) 0.0%(0) 100.0%(90) 注:括号内为相应的样本数。 -
[1] Shields J E, Karr M E, Tooman T P, et a1. The Whole Sky Imager—A Year of Progress. Eighth Atmospheric Radiation Measurement (ARM) Science Team Meeting, Tucson, Arizona, 1998. [2] Long C N, Slater D W, Tooman T. Total Sky Imager Model 880 Status and Testing Results. ARM technical Report ARM TR-006, US Department of Energy, Washington D C, 2001. [3] 吕达仁, 霍娟, 吕曜, 等. 地基全天空成像仪遥感的科学、技术问题和初步试验//童庆禧. 中国遥感——奋进创新20年. 北京: 气象出版社, 2001: 114-120. [4] Cazorla A, Olmo F J, Alados-Arboledas L. Development of a sky imager for cloud cover assessment. Journal of the Optical Society of America, 2008, 25(1):29-39. doi: 10.1364/JOSAA.25.000029 [5] 孙学金, 高太长, 霍东力, 等.基于非制冷红外焦平面阵列的全天空红外测云系统.红外与激光工程, 2008, 37(5): 761-764. http://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ200805006.htm [6] 张阳, 吕伟涛, 马颖, 等. 基于球瓣旋转遮光结构的地基全天空云自动观测系统. 中国专利: 200920277594. 5. 2009. [7] 谭涌波, 陶善昌, 吕伟涛, 等.双站数字摄像测量云高.应用气象学报, 2005, 16(5):629-637. doi: 10.11898/1001-7313.20050509 [8] 翁笃鸣, 韩爱梅.我国卫星总云量与地面总云量分布的对比分析.应用气象学报, 1998, 9(1):32-37. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=19980105&flag=1 [9] 高太长, 刘磊, 赵世军, 等.全天空测云技术及发展.应用气象学报, 2010, 21(1): 101-109. doi: 10.11898/1001-7313.20100114 [10] 杨俊, 吕伟涛, 马颖, 等.基于自适应阈值的地基云自动检测方法.应用气象学报, 2009, 20(6): 713-721. doi: 10.11898/1001-7313.20090609 [11] Yang J, Lu W, Ma Y, et al. An automated cirrus cloud detection method for ground-based cloud image. J Atmos Ocean Technol, 2012, 29:527-537. doi: 10.1175/JTECH-D-11-00002.1 [12] Huo J, Lu D. Cloud determination of all-sky images under low-visibility conditions. J Atmos Ocean Technol, 2009, 26: 2172-2181. doi: 10.1175/2009JTECHA1324.1 [13] 杨俊, 吕伟涛, 马颖, 等.基于局部阈值插值的地基云自动检测方法.气象学报, 2010, 68(6): 1007-1017. doi: 10.11676/qxxb2010.095 [14] 中国云图.北京:气象出版社, 2004. [15] Peura M, Visa A, Kostamo P. A New Approach to Land-based Cloud Classification. Proceedings of the Thirteenth International Conference on Pattern Recognition (ICPR'96), Vienna, Austria, 1996: 143-147. [16] Buch J K A, Sun C H. Cloud Classification Using Whole-sky Imager Data. 9th Symposium on Meteorology Observations and Instruments, Charlotte, North Carolina, 1995:353-358. [17] Singh M, Glennen M. Automated ground-based cloud recognition. Pattern Anal Applic, 2005(8): 258-271. doi: 10.1007/s10044-005-0007-5 [18] 孙学金, 刘磊, 高太长, 等.基于模糊纹理光谱的全天空红外图像云分类.应用气象学报, 2009, 20(2): 157-163. doi: 10.11898/1001-7313.20090204 [19] Calbó J, Sabburg J. Feature extraction from whole-sky ground-based images for cloud-type recognition. J Atmos Ocean Technol, 2008, 25: 3-14. doi: 10.1175/2007JTECHA959.1 [20] Heinle A, Macke A, Srivastav A. Automatic cloud classification of whole sky images. Atmospheric Measurement Techniques Discussions, 2010, 3: 269-299. doi: 10.5194/amtd-3-269-2010 [21] 孙君顶, 马媛媛.纹理特征研究综述.计算机系统应用, 2010, 19(6): 245-250. http://www.cnki.com.cn/Article/CJFDTOTAL-XTYY201006059.htm [22] Haralick R M, Dinstein I, Shanmugam K. Texture Features for Image Classification. IEEE Transactions on Systems, Man and Cybernetics, 1973: 610-621. [23] Tamura H, Mori S, Yamawaki T. Texture Features Corresponding to Visual Perception. IEEE Transactions on Systems, Man and Cybernetics, 1978: 460-473. [24] Stricker M, Orengo M. Similarity of Color Images. SPIE Storage and Retrieval for Image and Video Databases Ⅲ, 1995:381-392. [25] Hu M K. Visual Pattern Recognition by Moment Invariants. IEEE Trans on Information Theory, 1962: 170-179. [26] 冯伟兴, 唐墨, 贺波, 等. Visual C++数字图像处理模式识别技术详解.北京:机械工业出版社, 2010.