Research And Application Of Remote Sensing Scene Classification Algorithm Based On Deep Learning

Remote sensing scene classification is an important cornerstone of spatial analysis,from early methods based on pixels or objects to methods based on deep learning.Remote sensing scene classification can effectively obtain high-level semantic information to assist major remote sensing applications,which is of great significance to the development of land use monitoring,urban planning and precision agriculture.However,in the acquisition of remote sensing images,the high similarity between classes and the big difference within classes caused by the rapid and wide-ranging shooting of remote sensing images from different heights,angles and directions also bring a series of challenges to the subsequent classification tasks.Therefore,aiming at the problems of large intra-class differences and high similarity between classes in remote sensing data sets,this paper proposes a feature aggregation compensation convolutional neural network（FAC-CNN）based on bidirectional gating and active rotation aggregation of multi-scale features.and improved feature aggregation compensation-spatial pyramid pooling（FAC-SPP）for remote sensing scene classification.The main contents and achievements of this paper are as follows:（1）A feature aggregation compensation convolutional neural network FAC-CNN based on bidirectional gating and active rotation aggregation of multi-scale features is proposed.FAC-CNN combines features of different scales extracted from the branch network of multi-scale feature extraction part,and extracts shallow features with rich spatial information as part of the bottom convolution feature combination.And through bidirectional gated connection,it promotes the complementarity with the high-level features extracted by dense connection,and finally fuses the bottom,high-level and global features for classification.The experimental results show that the average recognition accuracy of FAC-CNN on NWPU-RESISC data set is improved by 2.05%and 2.69%,respectively,compared with Attention circular convolution Network（ARCNET）and Gated Bidirectional Network（GBNET）.On the aerial image dataset（AID）,the average recognition accuracy of FAC-CNN increased by 3.24% and 0.68%,respectively.It is verified that FAC-CNN can effectively use shallow spatial information to assist model classification.（2）A remote sensing scene classification model FAC-SPP based on double-layer spatial pyramid pooling is proposed to improve FAC-CNN.By controlling receptive fields,features of different scales can be effectively utilized,so that feature vectors containing rich spatial information can be extracted,and the expression ability of convolution features can be improved.The experimental results show that FAC-SPP improves 0.8%and 0.68%on NWPU-RESISC and AID datasets,respectively.FACSPP model based on hierarchical spatial pyramid pooling can effectively extract rotation scale invariant features,so that the model can better capture the spatial information in the scene,thus improving the accuracy of scene classification.（3）The remote sensing scene classification system is designed and deployed to the micro-service environment for application.Firstly,the farmland data set is constructed and the model is trained under the data set by using transfer learning to complete the farmland recognition scene.At the same time,FAC-CNN and FAC-SPP are applied to the remote sensing scene classification system,and finally the implementation and performance analysis of the system are completed.