HY-1C卫星CZI影像卤虫条带提取研究--以艾比湖为例  

Extraction of Artemia slicks from HY-1C CZI images:Taking Ebinur Lake as an example

作　　者：王欣 刘建强[3] 丁静[3] 田婧怡 孙相晗 田礼乔[1,2] WANG Xin;LIU Jianqiang;DING Jing;TIAN Jingyi;SUN Xianghan;TIAN Liqiao(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China;Hubei LuoJia Laboratory,Wuhan 430079,China;National Satellite Ocean Application Service,Ministry of Natural Resources,Beijing 100081,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,武汉430079 [2]湖北珞珈实验室,武汉430079 [3]自然资源部国家卫星海洋应用中心,北京100081

出　　处：《遥感学报》2023年第1期104-115,共12页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家重点研发计划(编号:2018YFB0504900,2018YFB0504904);国家自然科学基金(编号:42071325,42176183);新疆维吾尔自治区区域协同创新专项(编号:2020E01044);武汉大学测绘遥感信息工程国家重点实验室专项科研经费;武汉大学“985计划”;国家重点实验室装备专项基金。

摘　　要：基于HY-1C卫星海岸带成像仪CZI(Coastal Zone Imager)影像提取卤虫条带,对利用自主遥感数据开展生物资源监测与指导捕捞利用具有重要意义。本文以艾比湖为例,分析了HY-1C卫星CZI影像与Landsat-8卫星OLI数据的卤虫—水体端元光谱特征及差异;结合滑动窗口裁剪和光谱匹配因子SBAF(Spectral Band Adjustment Factors)模拟构建了有效样本量为837的卤虫—水体数据集;使用深度为5层的U型全卷积神经网络U-Net(U-Shaped Fully Convolutional Neural Network)算法提取卤虫条带并进行了评估与应用。与支持向量机法、最大似然分类法、归一化水体指数法相比,U-Net算法效率高、鲁棒性更好,卤虫条带的提取精确率和F1分数分别为92.02%和90.55%,比其他方法高出约11%—23%,即使面对复杂水体背景干扰,提取错误率也仅有3.3%;由2019年—2021年10景CZI影像的提取结果可知,研究期间卤虫条带的最大最小面积之比约为5.8,变化剧烈且与水体面积存在一定关联,但决定卤虫条带分布与面积的更多影响要素仍需进一步研究。未来,将建立多源、多分类、大样本量的遥感数据集,发展泛化能力更强的提取算法,实现长时序、大范围的盐湖卤虫条带时空规律分析。Artemia is a kind of small crustacean that lives in high salinity water, which can be used as an excellent fishery feed and an important component of carbon flux and biological chain in salt lakes. Because of its nonnegligible ecological and economic value, it is of great significance to develop a high-precision extraction method of Artemia based on remote sensing data for biological resource monitoring and reasonable fishing. Taking Ebinur Lake as an example, this paper proposed an automatic method to extract Artemia based on the HY-1C Coastal Zone Imager(CZI) images and deep learning technology. Firstly, the spectral characteristics of HY-1C CZI and Landsat-8 OLI sensors in the Artemia endmember were analyzed and the Spectral Band Adjustment Factors(SBAF) were used to eliminate the response differences between the two sensors to construct the Artemia-water dataset containing 837 effective samples of 64×64 size. Secondly, 70%of the dataset was used to train the U-Shaped Fully Convolutional Neural Network(U-Net) with a depth of 5, and the remaining 20% and 10% of the data were used to verify and test the algorithm, respectively. The model iterated 6700 times in the training process, which took 35 minutes. During this period, we used the adaptive moment estimation(Adam) optimizer with an initial learning rate of 1×10^(-4), and the binary cross entropy as the loss function. The training batch size was set to 4 since the equipment limitation. Whenever the loss value of the verification dataset did not decline within the last 3 epochs, the learning rate was halved. The training would be terminated automatically if it did not decline within the last 10 epochs. Finally, the impact factors and application potential of this method were further analyzed and discussed. The experimental results demonstrated that, compared with the Support Vector Machine(SVM), the Maximum likelihood Classification(MLC), and the Normalized Difference Water index(NDWI) algorithms, the extraction Precision and F1 score of U-Net were 92.02% an

关 键 词：HY-1C卫星 海岸带成像仪CZI 卤虫条带 艾比湖 U-Net 

分 类 号：P2[天文地球—测绘科学与技术]