Research On Ocean Front Detection Method Based On Multi-Scale Features

The ocean front,as an important mesoscale oceanic phenomenon,widely occurs worldwide and is often found in ocean basins.Since ocean fronts are at the intersection of water masses of different nature,accompanied by strong ocean mixing and stirring,enhanced biological productivity and ecological genetic crossover phenomena,they are key factors affecting sea-air interactions,ocean heat exchange and material transport.Therefore,achieving accurate detection of ocean fronts is an important foundation for analyzing the spatial and temporal changes of ocean fronts and monitoring the dynamics of ocean meteorology.With the rapid development of information technology and the continuous improvement of ocean observation technology,ocean science is entering the era of big data,and multi-source ocean observation data are being collected and stored at an unprecedented scale and speed.Satellite radiometers and Sea Surface Temperature（SST）inversion algorithms have been iteratively optimized,driving the accuracy and resolution of SST observations to increase dramatically,providing an accurate and abundant data source for ocean front studies.The current research methods on ocean front detection can be summarized into two categories:traditional ocean front edge extraction and automatic front detection based on deep learning.Traditional ocean front detection methods extract front features manually from the perspective of statistical analysis and rely on experts to set thresholds for front detection,and the methods are not universally applicable and the extraction effect is not ideal.Based on deep learning,ocean front detection methods can automatically extract front features through deep networks.However,ocean fronts have the characteristics of weak edges and small targets,and show obvious characteristic differences in different sea areas and seasons.How to enhance the feature expression ability of detection models to achieve accurate detection of ocean fronts is a challenging task.The objective of this research is to achieve accurate localization of ocean fronts and satisfy the detection requirements of different sea areas.Considering the limitations of existing methods and the critical influence of front feature extraction on detection performance,this research proposes an ocean front detection method based on multi-scale features.The main research contents and achievements of this research are as follows:（1）We construct an ocean front dataset based on Sea Surface Temperature（SST）.In this research,the SST data is provided by Remote Sensing Systems（REMSS）,which is the optimal interpolation infrared and microwave fusion data,and the Gulf of California and the Gulf of Mexico,where ocean fronts are concentrated and frequent,are selected as the study area,and the monthly average SST images of the two regions over 10 years（2010-2020）are used as raw data to build the required ocean front dataset.In this research,the dataset is manually labeled by manual visualization and expert adjudication,and the geometric transformation operation is used to expand the size of the dataset and ensure the diversity of the data.（2）We propose a multi-scale feature extraction-based ocean front detection method.The intersection of different water bodies in the ocean and slow temperature change lead to the characteristics of small targets and weak edges of ocean fronts.The existing detection methods have problems such as inaccurate portrayal of ocean fronts and false detection of image pixels.To address these problems,this paper proposes a multi-scale feature extraction-based detection method for ocean fronts.The method improves the detection ability of the model for edge contour and location information by designing a multi-scale feature extraction module,which obtains spatial and location features through a shallow learning network,while combining the high-level semantic features obtained by a deeper network;in addition,the method introduces a hybrid loss function DF_loss combined with Dice_loss and Focal_loss to guide the model to focus on the pixel-level differences between the predicted results and the labeled values to improve the accuracy of front pixel detection.The experimental results show that compared with the existing methods,the proposed method improves by 9%in IOU and F1 score,and achieves 89.98%in Recall,which can accurately locate the position and edge contour of the ocean front and carve out the accurate front morphology.（3）We propose a multi-scale feature enhancement-based ocean front detection Since the ocean fronts in different seas are dynamically changing,their generation and extinction cycles have obvious seasonal and sea area differences,and the fronts have different weight shares in different channels and spatial dimensions of the feature map,but the convolution operation fuses all channel information,which leads to the semantic information mixing among different channels.Therefore,in this research,based on the research（2）,the attention module of the convolution block is cascaded at the front and back ends of the multiscale feature extraction module to generate the attention feature mapping from the channel and spatial dimensions to enhance the detection performance of the model for different frontal features and spatial locations.A multiclassification ocean front dataset is also introduced in the model training to verify the generalization ability of the model.The experimental results show that the proposed method has advantages in both binary and multiclassification ocean front detection and can meet the different demands of ocean front detection.