Research On Sea Ice Detection Method Based On The Decomposition Of Mixed Pixels

Sea Ice Monitoring is of vital significance,not just because of the demand for emergency disaster prevention,but also its role in thermos-dynamic time-space feedback mechanism to support regional and global climate change analysis.In this paper,the data of Moderate-resolution Imaging Spectroradiometer(MODIS)with high spectral resolution,high temporal resolution,and moderate spatial resolution are selected as experimental data.And the Bohai Sea,one of the main ice-infested sea areas in China,is selected as the experimental area.Based on the characteristics of the MODIS data with a maximum spatial resolution of 250 m and the physical characteristics of the mixed spectral reflectance of sea ice and sea water,the decomposition of mixed pixels provides a feasible idea for sea ice detection.A Linear Spectral Unmixing method based on MultiConstraint Endmembers(LSU-MCE)is proposed in this paper.The method can be divided into two steps.The first is the extraction of endmembers based on multi-feature constraints.This is a rough classification process,it analyzes the differences between sea ice and sea water from four aspects: spectral features,texture features,shape features,and temperature features,and it selects multiple representative feature expressions to construct the rules,and then extracts constrained endmembers.The second is a fine classification process,it obtains the abundance by linear decomposition of mixed pixels,and then obtains the classification results of sea ice and sea water in sub-pixel level.The proposed method can automatically extract the endmembers of different classes.By constructing different feature constraint rules,a multi-constraint endmember purification method is proposed to make the refined endmembers more representative and further improve the accuracy of decomposition.There are two main innovations in the proposed method:(1)Considering the actual situation of high suspended sediment concentration in the experimental area,the turbid sea water is separated from the sea water as a separate endmember to analyze its characteristics.That is to say,three types of endmembers are extracted in the paper,namely “clear sea water endmember”,“turbid sea water endmember” and “sea ice endmember”.Turbid sea water has different characteristics from sea ice and clear sea water,and separating it can effectively solve the disturbance caused by suspended sediment.(2)The texture mean feature weighted by edge point density is proposed in the paper.It is difficult to distinguish between sea ice and turbid sea water based on the mean feature calculated directly from grey level co-occurrence matrix(GLCM).In this paper,the edge point density is used to weight the mean feature.The edge point density of the sea ice areas is large,and the mean feature of the sea ice can be effectively enhanced,while the edge point density of the turbid sea water areas is small,or even 0,and thus the characteristics of turbid sea water on the image can be effectively reduce,thereby increasing the separation of sea ice and turbid sea water.The paper carried out detailed experiments and fully analyzed the performance and robustness of the proposed method,and compared the results with the other three typical methods including the comparison of visual results and the quantitative evaluation based on the confusion matrix.The results show that the proposed method has higher overall accuracy and Kappa coefficient,and better visual results.And the method is less affected by stripe noise and thin clouds and can be applied to the images of Terra and Aqua.In addition,it has been tested on more images and all have good detection results.