陈琛, 马毅, 胡亚斌, 张靖宇. 一种自适应学习率的卷积神经网络模型及应用——以滨海湿地遥感分类为例[J]. 海洋环境科学, 2019, 38(4): 621-627. DOI: 10.12111/j.mes20190421
引用本文: 陈琛, 马毅, 胡亚斌, 张靖宇. 一种自适应学习率的卷积神经网络模型及应用——以滨海湿地遥感分类为例[J]. 海洋环境科学, 2019, 38(4): 621-627. DOI: 10.12111/j.mes20190421
CHEN Chen, MA Yi, HU Ya-bin, ZHANG Jing-yu. A Convolution neural network model with adaptive learning rate and its application-a case study of remote sensing classification of coastal wetland[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2019, 38(4): 621-627. DOI: 10.12111/j.mes20190421
Citation: CHEN Chen, MA Yi, HU Ya-bin, ZHANG Jing-yu. A Convolution neural network model with adaptive learning rate and its application-a case study of remote sensing classification of coastal wetland[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2019, 38(4): 621-627. DOI: 10.12111/j.mes20190421

一种自适应学习率的卷积神经网络模型及应用——以滨海湿地遥感分类为例

A Convolution neural network model with adaptive learning rate and its application-a case study of remote sensing classification of coastal wetland

  • 摘要: 滨海湿地是重要的生态系统,开展滨海湿地类型分布监测,对滨海湿地的保护与利用具有重要意义。传统卷积神经网络(CNN)模型中的学习率为人工设置的固定值,本文提出一种自适应学习率的CNN模型,以代价函数为目标函数自动计算学习率的优化值,从而使CNN模型具有自适应性。应用黄河口滨海湿地的CHRIS高光谱遥感影像数据,开展本文提出的CNN模型分类方法验证与优化。实验结果表明:对于不同的学习率搜索区间,自适应学习率CNN模型在0,1区间的整体分类精度最高,说明在学习率优化过程中只需在小区间0,1内进行微调就能保证较好的分类精度;对于不同的学习率初值,自适应学习率CNN模型的分类精度和稳定性都高于传统CNN模型,说明本文提出的模型对初值敏感性较低;在训练样本数目减少的情况下,两模型分类精度的稳定性都有不同程度的降低,但在保证训练样本占全部样本1.35%以上的条件下,自适应学习率CNN模型稳定性高,说明本文提出的模型对小样本具有一定的适应能力。

     

    Abstract: Coastal wetlands are important ecosystems.To monitor the wetland types distribution of the coastal wetland is of great significance for protecting and utilizing coastal wetlands.The learning rate in the traditional convolution neural network (CNN) model is a fixed value that is manually set.In this paper, a CNN model with adaptive learning rate is proposed, and the optimal value of the learning rate is automatically calculated by using the cost function as the objective function, which makes the CNN model adaptive.Based on the CHRIS hyperspectral remote sensing image data of the Yellow River estuary coastal wetland, the CNN model classification method proposed in this paper is validated and optimized.The experimental results show that:The adaptive learning rate CNN model has the highest overall classification accuracy in the interval of0, 1 for different learning rate search interval, which means that it is necessary to fine tune the interval0, 1 only to ensure better classification accuracy in the process of learning optimization.For the different the initial learning rate, the accuracy and stability of the adaptive learning rate CNN model are higher than those of the traditional CNN model, indicating that the model proposed in this paper is less sensitive to the initial value.In the case of reducing the number of training samples, the stability of the two model classification accuracy have different degrees of reduction.However, the adaptive learning rate CNN model has high stability under the condition that the training sample is more than 1.35% of the total sample, which indicates that the model proposed in this paper has a certain ability to adapt to the small sample.

     

/

返回文章
返回