Remote Sensing Imagery Scene Classification Based On Spiking Neural Network

With the success of deep learning technology,the availability of remote sensing image scene data and the increase of parallel computing resources,the algorithm based on deep neural network has become a leading algorithm in the field of remote sensing image research.However,as the resolution of remote sensing images is getting higher and higher,deep neural networks are becoming more and more complex,so the realization of remote sensing image scene classification tasks often requires more computing resources,which limits their practical applications.Compared with the deep neural network classification method,the remote sensing image scene classification method based on the spiking neural network is more biologically reasonable.This method has the characteristics of event-driven and low energy consumption,and can quickly and efficiently classify remote sensing images while maintaining performance comparable to deep neural networks.This paper is based on the theory of deep neural network,spiking neural network,etc.,by converting a trained deep neural network into a spiking neural network,a conversion algorithm is established to map the weight of the deep neural network to an equivalent spiking neural network,so as to narrow the performance gap between them,and build a spiking neural network model to classify remote sensing images.The main research contents of this paper are as follows:Firstly,because the previous remote sensing image feature extraction network such as VGG-16 models,not only has a large amount of parameters,but also cannot extract the features of remote sensing images well,so this paper proposes a deep neural network with low complexity and high classification accuracy.The model is used for conversion.It contains19 convolutional layers,1 global average pooling layer and 1 fully connected classification layer.Compared with the fine-tuned VGG-16(parameter 138M)and other models,the deep neural network model(parameter 20.9M)proposed in this paper has a higher overall accuracy.Secondly,in order to overcome the huge computational cost of deep neural networks,this paper adopts the idea of converting a trained deep neural network into a spiking neural network,and realizes a spiking neural network model based on conversion.Although the spiking neural network has achieved good results on the MNIST dataset and CIFAR10 dataset,its potential in remote sensing has not been studied and explored.The experimental results show that the conversion loss of this method between the models on the UCM dataset,WHU-RS dataset and RSSCN7 dataset is about 2%～10%,and the method has low power consumption.The results can be predicted after the first output spiking is generated.Finally,the channel normalization method is introduced to replace the previous layer normalization method,which solves the problem of insufficient spiking rate activation.On the basis of the original spiking neural network,a multi-bit spiking-based spiking neural network is proposed to reduce the layer-by-layer transmission of propagation errors.The experimental results show that the channel normalization method and the multi-bit spiking method can speed up the convergence speed of the conversion method and improve the accuracy of the conversion.Both the channel normalization method and the multi-bit spiking method have the best effect when applied together,which can realize the lossless conversion from deep neural network to spiking neural network on the UCM dataset,WHU-RS dataset and RSSCN7 dataset in less time steps.