Research On Recognition Of Sea Ice Region Based On MODIS Satellite Remote Sensing Image

Sea ice is a special phenomenon in high-latitude sea areas.When the ice condition is serious,it will cause great safety risks to the production and life of the country and people.MODIS data is an important resource for studying sea ice.Improving the segmentation accuracy of sea ice regions in MODIS data is of great significance to marine production and coastal economic activities in high latitude regions.The sea ice in MODIS data images is mostly irregular and regionally scattered.The traditional segmentation algorithm with artificial threshold adjustment has a low recognition rate for sea ice.The Unet network has less application in MODIS data images,and insufficient use of MODIS data information,resulting in lower accuracy of semantic segmentation results and inability to achieve better results.This paper improves and optimizes the classic Unet model for the sea ice region segmentation problem of MODIS data,and proposes a semantic segmentation network SIS-Unet(Sea Ice Segmentation Unet)for sea ice region recognition.The main tasks completed are as follows :(1)Exploring the preprocessing of MODIS data images and the production of data sets.The pre-processing of the image in the early stage includes correction and calibration,and then data labeling of the processed MODIS data.Due to the seasonal restriction of sea ice and other objective conditions,there are few sea ice images of MODIS data available at present.Therefore,data augmentation is carried out on the dataset.In order to facilitate the training and identification in the subsequent experiment,python programming is used to uniformly crop the annotated MODIS data set,which provides data support for the experiment.(2)This paper discusses the application of traditional threshold segmentation algorithms and deep learning networks to segmentation of sea ice in MODIS data images.To improve the classical semantic segmentation network(Unet),residual structure is added in the coding part.The residual module includes two different residual structures and performs multiple convolution operations.Compared with the traditional Unet model,it can extract deeper sea ice feature information while reducing the amount of calculation.Experiments show that the residual network can effectively identify the sea ice area in MODIS data,and obtain a more accurate sea ice area segmentation effect.(3)In order to extract the multi-scale information of MODIS data more effectively,this paper considers adding an Atrous Spatial Pyramid Pooling(ASPP)structure to the residual Unet modulel.By resampling the feature information of different scales in remote sensing images,the ASPP module effectively realizes the accurate segmentation of multi-scale sea ice area in MODIS data images,and integrates the multi-scale sea ice feature information while increasing the receptive field.Experiments have proved that ASPP structure is a complement to the convolution in Unet and enhances the characteristic information processing ability of the network.The SIS-Unet model proposed in this paper is superior to the traditional threshold method and the classic Unet model in many indicators,and can achieve a more accurate segmentation effect on the sea ice region in MODIS data.