Segmentation And Recognition Algorithm Of Cloud Organization Pattern In Satellite Image

With the wide application of satellite technology in the field of meteorology,cloud organization pattern in satellite image is widely used in the field of meteorological analysis and forecast.At present,the meteorological forecast based on cloud organization pattern in satellite images mainly relies on professional meteorological personnel to study and judge the cloud image data.Automatic and accurate segmentation and recognition of cloud organization pattern in satellite images has become one of the research hotspots in the field of meteorology and artificial intelligence.This thesis focuses on the segmentation and recognition algorithm of cloud organization pattern in satellite images based on deep learning,and explores effective segmentation and recognition strategies.With a large number of samples,deep learning algorithm uses deep neural network model to extract features from cloud organization pattern graph,and segmentation mask is generated and the category of cloud organization pattern is judged according to the extracted features.At present,the segmentation and recognition algorithm of cloud organization pattern based on deep learning is still in the exploratory stage.How to realize the automatic and accurate segmentation of cloud organization pattern is a research hotspot in the field of meteorology and artificial intelligence.In this thesis,we preprocess the dataset of cloud organization pattern,and apply pseudo-labelling technology to increase the number and diversity of samples.Then,the algorithm of UNet series is applied to the task of cloud organization pattern segmentation and recognition.The algorithm of UNet series is analyzed through experiments,and UNet is selected to segment and recognize the cloud organization pattern in satellite image.In order to improve the feature extraction ability of UNet,the EfficientNet structure is used as the encoder of UNet.In order to strengthen the supervision of model training,the recognition and supervision module is introduced in UNet,and the self-supervision attention mechanism is introduced in the recognition and supervision module to further strengthen the supervision of model training.In order to improve the ability of the decoder to filter and fuse different scale features,the decoder of UNet is stacked and pruned.The model improves the segmentation and recognition performance of cloud organization pattern in satellite image.Because the data used in training single model is fixed,the segmentation and recognition performance of single model has some limitations.In order to further improve the segmentation and recognition effect,the model integration strategy is designed and implemented.The cross-validation technology and Bagging method of model integration are applied to integrate multiple models obtained by using different methods or training on different datasets,so that each model on some data can be brought fully utilized.In the post-processing stage,the dilation and corrosion technology is used to process the mask,and the segmentation mask is pruned and screened to further improve the segmentation and recognition effect.The Dice score of quantitative cloud organization pattern segmentation and recognition is improved from 0.6419 to 0.6796.