HY-1C/DCZI多光谱影像云覆盖与耀斑区域绿潮自动提取方法研究

Automatic extraction of green tide in areas with clouds or solar flares in HY-1C/D CZI multispectral images

针对多光谱影像受云、雾、太阳耀斑等因素的影响,难以实现高精度的绿潮自动提取的问题,本文以我国的HY-1C/D卫星CZI载荷多光谱影像为数据源,采用数据挖掘技术,通过探索绿潮区域与非绿潮区域的光谱分布差异,提出一种适用于HY-1C/DCZI影像的高精度、全自动绿潮提取方法。首先,分析有云区域和无云区域样本的光谱差异,给出厚云去除规则;其次,选取绿潮和非绿潮区域的样本,采用决策树算法生成绿潮提取规则;然后,针对薄云和厚云边界区域常常会出现误检绿潮的问题,设计了5种错误类别修正策略。为验证方法的有效性,收集2021年黄海区域绿潮暴发周期内的25景HY-1C/D CZI影像,开展绿潮自动检测实验。结果表明,与传统的NDVI方法、VB-FAH方法等指数方法以及ResNet50、U-Net等深度学习方法相比,本文方法在准确度、Kappa系数、F1-Score和MIoU等指标上均优于其他方法,而且能够实现在厚云、薄云、无云、云斑和耀斑区域复杂情况下的绿潮的高精度自动提取。

Multispectral images are greatly affected by factors such as clouds,fog,and solar flares,which makes it difficult to automatically extract high-precision green tides under complex weather conditions.Based on the multi-spectral images of my country’s HY-1C/D satellite CZI payload,using data mining technology to explore the differ-ence in data distribution between green tide areas and non-green tide areas,we propose a high-precision and fully automatic green tide extraction method,which can be applied to HY-1C/D CZI sensor data.First of all,the thick cloud area is removed by preliminary extraction rules to achieve preliminary classification.Then,the correctly clas-sified green tide samples and non-green tide samples were used as positive and negative samples respectively,and these samples were used as experimental data to train the decision tree model,and the automatic extraction rules of green tide were obtained according to the model.Finally,5 strategies for correcting misclassifications were de-signed to achieve fully automatic extraction of green tides.In order to verify the effectiveness of the method,we collected 25 images of the green tide outbreak period in the Yellow Sea in 2021 for automatic detection experi-ments,and compared the experimental results with traditional index methods(NDVI,VB-FAH)and deep learning methods(ResNet50,U-Net).The results showed that the method outperformed other methods in terms of accuracy,Kappa coefficient,F1-Score,and MIoU.The accuracy of green tide extraction was higher in areas with thick clouds,thin clouds,cloudless clouds,cloud spots,and flares.