卷积神经网络在卫星遥感海冰图像分类中的应用探究——以渤海海冰为例

Application of convolutional neural networks in satellite remote sensing sea ice image classification:A case study of sea ice in the Bohai Sea

本文以TensorFlow为框架搭建卷积神经网络,基于迁移学习的思想,以经典的手写数字识别作为引入,对不同代价函数和激活函数组合对卷积神经网络模型分类结果影响进行了评价分析。以HJ-1A/B渤海海冰图像为实验数据源,分析了不同函数组合对遥感海冰图像分类的影响,优选出交叉熵代价函数与ReLU激活函数为最佳的组合,证明了卷积神经网络在遥感海冰分类中的应用可行性。对渤海海冰图像分类结果进行验证,其中带标签样本验证精度为98.4%。使用该模型对无标签的测试样本进行识别,讨论了样本的窗口尺寸对海冰分类结果的影响,发现在400×400小范围分类实验中最佳窗口尺寸为2×2;最后对整个渤海海域进行识别验证,效果较好。

This paper constructs a convolutional neural network based on TensorFlow.According to the idea of migration learning,the classical handwritten digit recognition is introduced as an introduction.The influence of different cost functions and activation function combinations on the classification results of convolutional neural network models is evaluated.Taking HJ-1A/B sea ice images as experimental data source,we analysis the influence of different function combinations on remote sensing sea ice image classification.It turns out that the cross-entropy cost function and the ReLU activation function are optimally combined.The feasibility of CNN in remote sensing sea ice classification is proved,and the classification results of the sea ice images in the Bohai Sea are verified.The calibration accuracy of the labeled samples is 98.4%.The model is then used to identify the unlabeled test samples.The influence of the window size on the sea ice classification results is discussed,and the optimal window size is 2×2 in the 400×400 small-scale classification experiment.Finally,the identification and verification of the entire Bohai Sea area is carried out,and the effect is good.