云是一种重要的天气现象，可靠的云观测技术对气候研究、天气分析以及天气预报等工作具有重要的意义。地基云观测是一种重要的云观测方式，它能够反应云的微观结构信息，弥补卫星观测的不足，充分应用地基云观测信息能够为云观测相关应用提供更全面的数据。在地基云观测中，地基云图自动分类技术是实现地基云自动观测的关键，该技术的应用不仅可以将观测员从繁重的云状观测工作中解放出来，而且可以提高云观测的准确性和时效性，因此实现地基云图自动分类具有重要意义。 目前，地基云图分类的相关研究工作尚处于起步阶段，尚有许多问题亟待解决，因此，本文将针对地基云图分类课题开展研究。由于地基云图属于自然纹理图像，本文将基于纹理图像分类方法，并结合地基云图本身的特征，对地基云图分类关键技术开展研究。本文的研究工作主要包括以下内容： 1. 针对地基云图自身纹理结构的复杂性，提出基于显著性局部二值模式（Salient Local Binary Pattern，SLBP）的地基云图分类方法。 该方法针对局部二值模式（LBP）描述子中一致性（uniform）模式并非地基云图中最主要模式这一问题， 提出显著性局部二值模式描述子。该描述子利用最频繁出现的模式获取地基云图中的显著性信息， 使提取的特征更具判别性，同时对噪声具有更好的鲁棒性。 2. 针对地基云图中普遍存在的光照变化，提出光照不变完备化局部三值模式（Illumination-invariant Completed Local Ternary Pattern ，ICLTP）描述子，并应用二次卡方距离计算两个直方图的相似性。 提出的ICLTP描述子通过引入光照不变子来克服光照变化；同时，通过采用三值编码策略表示更丰富的云图信息，并能够克服小幅噪声的影响。 3. 针对传统稀疏编码和最大化抽取操作相结合时容易导致信息丢失的问题，提出基于自适应符号稀疏编码的地基云图分类方法。 该方法充分考虑稀疏编码系数的符号对最大化抽取的影响，将约束稀疏编码系数符号的正则项加入到传统的稀疏编码模型中，从而有效的调节非零稀疏编码系数的正负比例，使得尽可能多的非零稀疏编码系数为正值。该算法不仅可以减少重构误差，还考虑了编码系数符号对最大化抽取的影响。 4. 针对单次随机映射在特征降维时的不稳定性，提出利用多次随机映射算法（Multiple Random Projections ，MRP）对云图进行分类。该方法通过融合多次随机映射的结果获得一个优化的码本， 利用该码本可以获得比单次随机映射更有判别性的特征。 5. 大部分地基云图分类方法均以单张地基云图为研究对象，实际上地基全天空成像仪器是按照设定的频率获取地基云图样本， 因此得到的数据为云图序列。相邻的地基云图之间包含对分类有用的时间信息，而现存的基于单张地基云图分类的方法却忽略了云图序列中相邻云图之间的时间信息。为了克服以上缺点，本文提出基于张量集合的地基云图序列模型（Tensor Ensemble of Ground-based Cloud Sequences，eTGCS），该模型不仅可以对地基云图序列进行建模、分类，还可以根据得到的模型参数合成新的地基云图序列。
英文摘要
Clouds are one of the most important weather phenomena and reliable cloud observation technology is of great significance for climate research, weather analysis and forecasting. Ground-based cloud observation, which can reflect the microstructure information of clouds and make up the disadvantages of satellite observations, is an important way of cloud observation. Thus, more comprehensive information can be obtained when the data of ground-based cloud observation is fully used. In the field of ground-based cloud observations, automatic ground-based cloud classification technology is the key technology to achieve automatic observation for ground-based cloud. The application of this technology can not only relief the pre-trained observers from huge manual cost, but also improve the accuracy and real-time performance of cloud observations. Hence, automatic classification for ground-based cloud is very important. At present, the related research of ground-based cloud classification is still in its infancy, and there are still many problems to be addressed. Therefore, this thesis focuses on ground-based cloud classification. Since ground-based cloud images are one kind of natural texture images, we will study the key technology of ground-based cloud classification based on the methods of texture image classification combined with the characteristics of the ground-based images. Main content of this thesis can be concluded as follows: 1. For the texture structural complexity of ground-based images, a novel ground-based cloud classification method based on salient local binary pattern (SLBP) is proposed. In order to solve the issue that the uniform LBP is not the most important pattern in the ground-based cloud images, salient local binary pattern (SLBP) descriptor is presented. The proposed descriptor takes advantage of the most frequently occurred patterns to capture descriptive information of ground-based cloud images, which makes the representation more discriminative. In addition, the proposed descriptor is more robust to noise. 2. In order to overcome the vast illumination changes in ground-based cloud images, the illumination-invariant completed local ternary pattern (ICLTP) is proposed. In addition, Quadratic-Chi metric is applied to calculate the similarity between two histograms. The proposed descriptor can conquer the illumination changes by introducing an illumination invariant factor; meanwhile, by employing the 3-valued coding strategy, the pr...
修改评论