Recognition Of Convective Area Based On Satellite Cloud Image

Disasters caused by severe convective weather have brought huge losses to the safety of people’s lives and property and national infrastructure.Accurate identification of strong convective clouds is an important prerequisite for effective prevention of severe convective disasters.The main work of this thesis revolves around the three infrared channel data of the FY-2G geostationary satellite and the radar combined reflectivity data.Aiming at the identification of strong convection areas in the satellite cloud image,this thesis combines computer vision technology and deep learning algorithms to design a set of identification A semantic segmentation algorithm for strong convective regions of satellite cloud images,and the algorithm is enhanced by machine learning.main tasks as follows:(1)This thesis collects the satellite cloud image data and radar reflectivity data required for this study,and on this basis,studies how to align the two types of images with temporal and spatial resolution.Later,the satellite cloud image is used as the original image,and the radar image is used as the annotation to construct a semantic segmentation data set.(2)Aiming at the problem of identifying convective areas in satellite cloud images,this thesis designs an improved PSPNet semantic segmentation algorithm,using the Res Net50 deep neural network as a feature extraction network.Aiming at the small and uneven distribution of strong convective areas,the pyramid pooling module is improved,and the fusion is More small-scale feature information makes the model’s receptive field focus on small areas,allowing the model to acquire more semantic information of these small areas.The final recognition accuracy rate is increased to79.24%,and the intersection ratio is increased to 59.85%.(3)The idea of using a machine learning model to improve the performance of the semantic segmentation model is realized.In this thesis,the satellite cloud image is divided into superpixel regions by SLIC and 15-dimensional feature space is extracted.The Bagging-SVM algorithm is designed to reduce the number of non-strong convective superpixels.The mask is made based on the recognition results of the algorithm,and the test set recognition result map of the semantic segmentation model is multiplied by the mask,and finally the intersection ratio is increased to 60.38%,which enhances the effect of the semantic segmentation model in this thesis.