| Weather has been closely watched since the beginning of human productive life,and more accurate prediction of weather allows humans to gain greater initiative in adaptation and confrontation with nature.The destructive power of extreme weather is immeasurable,and typhoons are one of the most dangerous and long-lasting types of extreme weather.With the rapid rise of neural network and deep learning technologies,the field of computer vision has also developed rapidly and become one of the most widely used and mature subfields of artificial intelligence technology.Object detection,as one of the most important subtasks in computer vision,is also widely used in various fields of production life.Applying object detection to the task of typhoon detection on satellite cloud images can greatly reduce the labor cost of typhoon forecasting.The research in this paper addresses three problems that arise for the task of typhoon object detection in satellite images.First,for the typhoon detection task,an object detection dataset for satellite clouds is constructed by using satellite clouds with historical typhoon forecast information with human-assisted adjustment in the absence of relevant datasets.Secondly,for typhoon samples with low wind speed and small sizes,a multi-stage detection algorithm is proposed to improve the detection capability of the model for low wind speed and small size samples by first screening the possible locations and then finely locating the optimized object boxes.Finally,for the presence of multi-source satellite images in the dataset,a contrastive learning algorithm based on multi-source satellite inputs is proposed to enhance the feature extraction capability of the model by using the object boxes in the labeled information to form positive and negative sample pairs,build connections between the local information and the global information,so that the detection performance of the model for typhoon objects can be improved.Based on the above research,an automatic typhoon detection system for satellite images is established,which will automatically detect typhoon objects from satellite images and save their predict results for viewing. |
