Research On Cloud Type Recognition In Ground-based Cloud Images

With the development of observation equipments and imaging technologies,observating and recording by human has been gradually replaced by imaging devices.But the interpretation of ground-based cloud image still depends on professional observers.This is due to that most of the current automatic classification and recognition methods of groundbased cloud images can not effectively recognize clouds as defined in meteorological observation standards.They can only recognize some typical clouds or cloud types according to other simplified classification criteria.Therefore,this thesis proposes an effective method for cloud type recognition according to the definition of cloud types in meteorological observation standards,and implements an automatic observation system based on ground-based cloud images.Cloud type recognition based on Ground-based cloud images can be regarded as an image classification problem.Feature extraction plays an important role in it.In this thesis,different feature extraction methods are used to represent the ground-based clouds.It is found that the low-level features such as color,texture and structure complement each other.The high-level features based on CNN can also work.Then,a cloud recognition method based on multi-view feature fusion is proposed,and a local feature edcoding mechanism are used in this method.The proposed method can classify the ground-based cloud images in a standard and hard classification principle.In order to further improve the cloud recognition in ground-based cloud images,a method which extracts multi-level and multi-scale local features based on pre-training CNN model is proposed,and Fisher Vector is proved to be better than full connection layer of CNN for encoding local features.What's more,it is found that there are many redundancies in the local patterns obtained by dense local sampling.These redundant local patterns do not help or even have a negative impact on the cloud classification.Therefore,a clustering-based local pattern mining method is proposed and applied before local feature encoding,and this makes the final extracted global features more discriminative.At present,most of ground-based cloud recognition methods are independent of cloud coverage estimation.To solve this problem,a supervised learning-based semantic segmentation method for all-sky images is proposed.Firstly,the whole-sky images are segmented into super-pixels which can keep the edges.Then feature extraction and cloud classification are carried out for these super-pixels.In addition,a feature space transformation method based on class-specific metric learning and subspace alignment is proposed,which defines a specific feature space for each type of cloud so that it is more discriminative from other types.Semantic segmentation of all-sky image can not only get the cloud category information in the sky,but also get the coverage and distribution of all kinds of clouds in the sky.Finally,in order to apply cloud recognition technology to actual meteorological observation,a multi-task automatic ground cloud observation system based on all-sky image is implemented.In this system,an image distortion correction method for all-sky image is proposed as image preprocessing.On the other hand,a multi-task fusion decision for all-sky image is made.In this system the informations including cloud types and coverage estimation are obtained simultaneously.Compared with the results of human observors,the automatic observation system proposed can achieve comparable or even better results.