Research On Typhoon Object Detection And Intensity Estimation Methods Based On Temporal Remote Sensing Cloud Images

Typhoons are a common natural weather phenomenon that have a serious impact on the social order and normal lives of residents in coastal areas.With the continuous development of the economy,the population density in China’s coastal areas continues to increase,and the losses caused by typhoon activities each year also continue to rise.Therefore,accurate monitoring and control of typhoon activity throughout its lifecycle,and accurate prediction of typhoon disasters,are of great significance for the safety of the lives and property of the people in China’s coastal areas.The use of meteorological satellite remote sensing cloud image data to locate typhoon targets and estimate typhoon intensity is a very effective method.However,the current mainstream methods of typhoon positioning and intensity determination are subjective,and errors that occur during the typhoon analysis process will continue to accumulate.The use of deep learning technology to locate or determine the intensity of typhoons is still not mature.Therefore,this article explores how to combine deep learning technology with typhoon positioning and typhoon intensity determination tasks to achieve a more objective and accurate typhoon forecasting plan.This article proposes the use of deep learning object detection algorithms to recognize typhoon targets,addressing the issue of subjectivity in traditional typhoon positioning methods.The characteristics of HMW-8 meteorological satellite data are studied,and its satellite remote sensing data is converted into a visual grayscale image and organized into a typhoon target detection remote sensing dataset.The improved FPN+Faster R-CNN object detection algorithm is used to detect unknown typhoon targets.To address the issue of high false detection rates in the target detection results,this article proposes a temporal post-processing algorithm.By establishing links between typhoon target detection results in adjacent time periods,erroneous detection targets are effectively screened out.Relevant experiments conducted on the typhoon target detection remote sensing dataset have verified the feasibility and effectiveness of this article’s method.This article addresses the issue of uneven feature distribution in the feature space of typhoon intensity determination tasks by proposing a category-instance joint feature learning framework.Classical metric learning loss functions are studied,and an IDC loss function based on metric learning is proposed to reduce the error caused by model classification errors by learning a feature space where the feature distance is proportional to the label distance.In addition,the cross-entropy loss function and IDC loss function are combined to further reduce the intra-class distance and improve classification accuracy.Furthermore,in view of the continuous trend of typhoon intensity changes over time,this article proposes a step smoothing algorithm,which greatly reduces the error in intensity estimation with almost no additional parameters or additional computational complexity.Relevant experiments conducted on the TCIR dataset have verified the feasibility and effectiveness of this article’s method.