遥感图像能够更为丰富地反映地物内部信息,但多种应用场景下会出现时间序列关键特征难以提取,导致图像分类效果不佳的问题.为此,设计一种面向多应用场景的遥感图像长短期记忆(Long Short Term Memory,LSTM)网络分类模型.对场景遥感图像时间序列特征进行表示;改进动态时间弯曲质平均算法,利用改进算法提取场景遥感图像时间序列关键形态特征;构建LSTM模型,利用梯度下降法训练LSTM模型,更新网络的权值与偏置,在训练完成的网络内,输入时间序列关键形态特征,输出遥感图像场景分类结果.实验结果表明,所设计模型可有效分类遥感图像场景,对场景遥感图像分类的Kappa值为0.97,分类耗时5.6 s,分类不同类别遥感图像场景时的预测分布方差最大为0.6,该方法的分类精度较高,且消融实验结果显示所设计模型的召回率高达95％,F1值高达0.96,由此可见,所设计模型对场景遥感图像分类具有显著的有效性.
LSTM Classification Model of Remote Sensing Images for Multi-application Scenes
Remote sensing images provide rich internal information of the ground,but it is often difficult to extract the key features of time series in various application scenes,resulting in poor image classification effect.A Long Short Term Memory(LSTM)classification model for multi-application scenes is designed.Firstly,the time series features of scene remote sensing image are represented,then the dynamic time bending quality averaging algorithm is improved to extract the key morphological features of scene remote sensing image time series,finally,the LSTM model is constructed,then trained by gradient descent method,the weights and bias of the network are updated,the key morphological features of time series are input in the trained network,and the scene classification results of remote sensing images are output.Experimental results show that the designed model can effectively classify remote sensing image scenes,with Kappa value of 0.97 and classification time of 5.6 s.The max variance of prediction distribution of different scenes is 0.6,namely the classification accuracy of the method is high,and the ablation experiments show that the recall rate and Fl value are 95％and 0.96,respectively.Thus,the designed model has significant effectiveness on scene remote sensing image classification.
国家科技期刊平台
NSTL
万方数据
