Classification Of Remote Sensing Scenes Based On Convolutional Neural Network

A growing number of remote sensing scene images with high complexity and information have been generated with the development of remote sensing technology in recent years.However,the existence of redundant background information and complex scale spatial distribution within the images,which are combined with the problems of intra-class diversity and inter-class similarity within the scenes,have led to the accurate and efficient classification of images becoming a research hotspot and difficulty in the field of remote sensing scene classification.The conventional image classification methods of remote sensing scene at pixel level or object-oriented remote sensing scene image classification can no longer meet the current classification needs.The exploration of the connection between the groundlevel features of images to the semantic information of scenes to obtain more accurate classification results has become the focus of current research.A new solution idea for remote sensing scene image classification task has been provided in recent years with the development of convolutional neural networks.It has been a new solution for the task of classifying remote sensing scene images.However,the existing models do not pay enough attention to the deep multiscale features and the features of key semantic objects in remote sensing scene images.This paper investigates the problems existing convolutional neural networks in the application of remote sensing scene image classification.The components of the study are as follows.(1)A MSRes-Split Net model for remote sensing scene image classification based on multiscale features and attention mechanism is proposed.First,we extract multiscale features by constructing a MSRes extraction module.Then,the multi-channel local feature information is weighted by Split-Attention block.We then fuse the global and local information by convolution operation to obtain the multi-scale feature representation of remote sensing scene images.Finally,the proposed method is experimentally validated on four publicly available datasets and compared with other state-of-the-art methods.The experimental results show that the method MSResSplit Net proposed in this paper has good performance.(2)The LKNet-RS model based on large-size convolutional kernel for remote sensing scene image classification is proposed.Firstly,by building a deep convolution module based on large-size convolution kernels,large-scale features are extracted from the scene images with complex and dense spatial distribution.Secondly,the key local features captured by the small-sized convolution kernel are fused with the large-scale features using structural reparameterization.Then reuse of the input features is achieved using shortcut residual join.Finally,the LKNet-RS model is experimentally validated on four datasets,and the results show that the model has superior classification performance.