一种稀疏降噪自编码神经网络影像变化检测方法

采用粤西2018—2019年优于0.5 m地理国情监测影像,结合2018年矢量化地表覆盖分类成果,使用稀疏降噪自编码神经网络深度学习方法,首先选取BJ、GF、ZY等各类高分辨率遥感影像训练生成多传感器训练模型;其次利用PCA主成分分析提取样本数据最大特征,实现样本数据白化降维;最后采用tanh函数作为神经元激活函数,选...

A Method of Sparse Noise Reduction Self-encoding Neural Network for Image Change Detection

Combined with the vectorized land cover classification results in 2018,this paper uses the better than 0.5 m geographic national conditions monitoring images from 2018 to 2019 in Western Guangdong.It uses the sparse noise reduction self-encoding neural network deep learning method to first select BJ/GF/ZY and other types of high resolution remote sensing images to generate a multi-sensor training model.Secondly,Principal Component Analysis(PCA)is used to extract the largest features of the sample data to achieve whitening and dimensionality reduction of the sample data.Finally,the tanh function is used as the neuron activation function,and the softmax regression classifier is selected to train the deep network layer by layer,calculate the overall loss function of the model,and find the minimum value of the loss function as the optimal parameter solution of the model.Experiments show that:1)This method has a recall rate of better than 88%in the urban-rural fringe,better than 92%in built-up areas,better than 80%in rural areas,and better than 68%in accuracy.2)Combined with semantic segmentation,the sparse noise reduction self-encoding neural network can produce better local extrema,and the network structure achieves better generalization performance,reducing the over-fitting phenomenon of high-resolution remote sensing image features.