Solid hydrometeor classification and riming degree estimation from pictures collected with a Multi-Angle Snowflake Camera
Solid hydrometeor classification and riming degree estimation from pictures collected with a Multi-Angle Snowflake Camera
复制标题
根据多角度雪花相机采集的图片进行固体水凝物分类和雾化程度估计
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
A. Berne
中科院分区：
文献类型：
--
作者：
C. Praz;Y. Roulet;A. Berne
抽象的。提出了一种基于多角度雪花相机（MASC）图像的固体降水自动分类方法。对于每个单独的图像，该方法依赖于一组基于几何和纹理的描述符的计算，以同时识别水凝物类型（在六个预定义的类别中），估计霜化程度和检测融化的雪。分类任务是通过一个正则化的多项式逻辑回归（MLR）模型训练超过3000 MASC图像手动标记的视觉检查。在第二步中，由MLR提供的概率信息在MASC的三个立体视图上进行加权，以便为每个水凝物分配唯一的标签。所提出的算法的准确性和鲁棒性进行评估，在瑞士阿尔卑斯山和南极洲收集的数据。该算法取得了很高的性能，与水流星类型的分类准确率和Heidke技能得分分别为95%和0.93。通过引入一个范围在零（无霜）和一（霰）之间的霜化指数来评价霜化程度，其特征在于可能的误差为5.5%。通过与现有的基于二维视频光盘（2DVD）数据的分类方法的比较进行了验证研究，表明这两种方法是一致的。
Abstract. A new method to automatically classify solid hydrometeors based on Multi-Angle Snowflake Camera (MASC) images is presented. For each individual image, the method relies on the calculation of a set of geometric and texture-based descriptors to simultaneously identify the hydrometeor type (among six predefined classes), estimate the degree of riming and detect melting snow. The classification tasks are achieved by means of a regularized multinomial logistic regression (MLR) model trained over more than 3000 MASC images manually labeled by visual inspection. In a second step, the probabilistic information provided by the MLR is weighed on the three stereoscopic views of the MASC in order to assign a unique label to each hydrometeor. The accuracy and robustness of the proposed algorithm is evaluated on data collected in the Swiss Alps and in Antarctica. The algorithm achieves high performance, with a hydrometeor-type classification accuracy and Heidke skill score of 95 % and 0.93, respectively. The degree of riming is evaluated by introducing a riming index ranging between zero (no riming) and one (graupel) and characterized by a probable error of 5.5 %. A validation study is conducted through a comparison with an existing classification method based on two-dimensional video disdrometer (2DVD) data and shows that the two methods are consistent.