基于密集连接空洞卷积神经网络的青藏地区云雪图像分类

为了提高高纬度地区云雪卫星图像的识别准确率,提出了密集连接空洞卷积神经网络与空洞卷积相结合的方法进行云雪卫星图像识别研究。该方法首先采用常规卷积层对图像进行处理得到特征图,然后采用多个密集块和过渡层对特征图进行处理。其中,密集块中采用跨层连接的方式实现了网络中所用层的特征传递,使得大量云雪特征得到重用,同时减轻了训练过程中的梯度消失问题。密集块中的卷积核采用空洞卷积,在减少参数量的同时扩大局部感受野,对云雪的光谱信息进行特征提取。最后,该方法采用平均全局池化层与全连接层得到云雪图像的预测结果。实验结果表明,与其他机器学习方法相比,该方法能够提高卫星云雪图像的识别准确率,具有良好的泛化能力。

Cloud and Snow Image Detection in Qinghai-Tibet Area based on Dense Dilated Convolution Neural Network

In order to improve the recognition accuracy of cloud-snow satellite imagery in Qinghai-Tibet region, this paper proposes a method combining dense dilated convolutional neural network and dilated convolution to carry out cloud snow satellite image recognition research. The method firstly processes the image by using a conventional convolution layer to obtain a feature map, and then uses a plurality of dense blocks and a transition layer to process the feature map. Among them, the feature transfer of the layers used in the network is realized by using the cross-layer connection in the dense block, so that a large number of cloud snow features are reused, and the gradient disappearance problem during the training process is alleviated. The convolution kernel in the dense block adopts the dilated convolution to expand the local receptive field while reducing the parameter quantity, and extract the feature information of the cloud snow. Finally, the method uses the average global pooling layer and the fully connected layer to obtain the prediction results of the cloud snow image. The experimental results show that compared with other machine learning methods, this method can improve the recognition accuracy of satellite cloud image and has good generalization ability.