Research On Segmentation Of Convection Cloud Based On Satellite Image Sequence

Disastrous weather such as typhoons,heavy rainfall,and thunderstorms often cause adverse effects on people's production and life,not only causing economic losses,but also very likely to threaten people's lives.If the entire life cycle of convection cloud can be accurately attained in advance,early warnings can be issued in time to reduce the impact of such disasters.This dissertation realizes the convective weather prediction based on the brightness temperature image sequences captured by the domestic Feng Yun-4 geostationary satellite,which can be divided into two steps: firstly,predicting the subsequent image sequences based on the satellite brightness temperature image sequences,and then segmenting the convection cloud based on the predicted brightness temperature image sequences.Therefore,the research content in this dissertation can be divided into two sub-tasks of image sequence prediction and semantic segmentation,which can also be regarded as a future scene parsing task.This dissertation needs to solve two coupled problems of image sequence prediction and semantic segmentation,but no algorithm has been found for segmenting the predicted satellite image sequence.For image sequence prediction task,most current models predict the target sequence from the input sequence based on the encoder-decoder structure.Due to the uncertainty of the real world,such models tends to make the predicted image farther away from the input sequence more ambiguous than the one close to it,and the complexity of the models is usually high,which is not suitable for large-scale image sequences.For image segmentation tasks,the existing models are mainly based on an input image,and the objects in the image are mostly rigid bodies,which do not take advantage of the temporal information in the sequence data and the physical characteristics of convection cloud.For image sequence segmentation task,most models aim to improve the segmentation efficiency of the image sequence by using the temporal information rather than improving the segmentation performance,so they are not applicable to the data which is sparse in time dimension.In the future scene parsing problem,there are few relevant researches,and the existing models only perform segmentation on single image obtained by single-step prediction based on the full-color image sequence,but not segment the image sequence obtained by multi-step prediction.Based on the spatio-temporally coupled model,this dissertation address three problems: difficulty in adjusting the loss weights of the multitasking model,insufficient extraction of short-term temporal information,and adverse effect of fuzzy predicted image.A multitasking weight dynamic adjustment module is constructed to accelerate the model convergence,and a Three-Dimensional Convolutional Neural Network with Gated Recurrent Unit(3D CNN-GRU)module is built to enhance the short-term feature extraction and a shape feature extractor is constructed to alleviate the impact of the fuzzy prediction image on the segmentation.The critical success index of our model in this dissertation is improved by about 5% compared to U-Net and by about 2% compared to 3D CNN with Long Short-Term Memory(3D CNN-LSTM)and Spatio-Temporally Coupled Generative Adversarial Networks(STC-GAN),where the improvement brought by the 3D CNN-GRU module is mainly reflected in the images with small prediction steps,while the improvement brought by the shape feature extractor is mainly reflected in the images with large prediction steps.