High-resolution Remote Sensing Image Scene Classification Based On Convolutional Neural Network

In recent years,China has built a comprehensive transportation spatiotemporal big data management platform and industry analysis system based on high-resolution remote sensing images.This system can improve the comprehensive management of transportation planning,construction,operation,development and service quality.With the advent of the era of remote sensing big data,the quantity and quality of remote sensing data have been greatly improved.Therefore,data-driven deep learning algorithms have been widely used and achieved remarkable achievements.Compared with images of natural scenes,remote sensing images are characterized by broad picture size,complex target elements and diverse image modal types.Therefore,more complex deep learning algorithm models are needed to capture image features to complete corresponding interpretation tasks.As one of the foundations of intelligent interpretation of high-resolution remote sensing images,there are two main difficulties in scene classification tasks:First,how to use deep learning models to extract complex feature information from massive amounts of remote sensing data.Make the model get better feature representation ability and improve the accuracy of model interpretation.Second,as deep learning models become more and more complex,the model requires high computing power resource support.Therefore,the edge end needs to resolve the contradiction between model complexity and power consumption.Practical applications have the problem of compressing and accelerating convolutional neural networks to further reduce the difficulty of deployment on edge computing devices.In response to the above difficulties,the research content of this paper mainly includes two points:(1)High-resolution remote sensing images have a series of characteristics such as multiscale,rich features,a large number of complex backgrounds and small targets.Therefore,this paper improves the traditional convolutional network based on the representation ability of the optimized feature extractor.The model is combined with the Gabor convolution layer to form an end-to-end network with parameter self-learning ability.At the same time,the high-order global covariance pooling is used to replace the low-order global average pooling in the traditional convolutional networks.This method improves the representation learning ability of the network and the generalization ability in the data set.The experiment was tested on four different remote sensing public datasets.Experimental results show that this network is superior to other existing networks in the classification of high-resolution remote sensing images.(2)With the continuous development of deep learning theory,the depth and complexity of traditional convolutional neural network models are also increasing.The parameters and calculations of the model are also increasing rapidly with the improvement of network performance.Therefore,many models are difficult to deploy in actual edge computing devices.This phenomenon limits the application of artificial intelligence technology in actual remote sensing intelligent interpretation.In this paper,an end-to-end lightweight convolutional neural network is designed for the above problems.The model uses a deep residual network as the backbone.This paper designs a lightweight convolution module based on the idea of feature reuse.At the same time,the spatial pyramid channel attention mechanism is integrated into the model structure.The model was tested on four different public remote sensing datasets.The experimental results prove that this method can greatly reduce the model parameters and the amount of calculation without losing the accuracy of the model.