Scene Classification Of High-Resolution Remote Sensing Image Based On Multi-features Deep Learning

Remote sensing image scene classification of complex surface has always been a research hotspot in the field of remote sensing image analysis.High-resolution remote sensing image is an important data source for ground object recognition.In the early days,the classification of remote sensing scenes was generally completed by visual interpretation of full-time personnel.The classification of complex large-scale remote sensing scenes required a lot of manpower,financial resources and time.With the rapid development of computer and aerospace technology,rich high-resolution remote sensing images provide researchers with a large number of data sources for Earth Observation(EO).Due to the sheer volume of data and the certain subjectivity and limitations of human interpretation,scientists began to seek automated methods to interpret images.Scientists have proposed physical indicators based on underlying visual features and their combinations,such as building volume,density,and alignment.Although high-resolution remote sensing images can clearly reflect the ground object information,problems such as illumination and sensor angle lead to large differences in scene types and high similarity between classes,which increases the difficulty of automatic interpretation of remote sensing images.When entering a more elaborate scene,this traditional pixel-level bottom-level feature analysis method cannot be directly migrated to other image scenes.If you do not use automated methods,updating such a database will be very resource intensive.Then there is object-oriented remote sensing image scene classification,but image preprocessing is needed.When the amount of data is large,the method has a large workload and consumes a lot of time,which is inconvenient for migration learning and poor processing results for image details.In recent years,deep learning has broken through the constraints of traditional visual algorithm structure,bringing new solutions for remote sensing image processing.At the same time,the classification accuracy of the high-resolution remote sensing image scene and the reuse rate of the code are improved,and it is easier to migrate learning.Thereby,deep learning promotes the intelligentdevelopment of remote sensing images.However,the network model usually only has the input of one network layer as the output of its upper layer.The deeper the network model,the lower the resolution of the feature map of the final classification.When the network model is shallow,the model learning ability is weak.The classical network model generally only considers the high-level features of the image,ignoring the characteristics of the bottom layer of the image,which leads to low utilization of image features,poor generalization ability of the model and low accuracy of scene classification.Therefore,this paper proposes a high-resolution remote sensing image scene classification method based on multi-level features for deep learning,which mainly carries out the following work:(1)This article will introduce the basic theories and methods of deep learning and scene classification,including several major frameworks and major networks for deep learning.Several major network structures are also elaborated,such as neural networks,convolutional neural networks and full convolutional neural networks.The related knowledge of high-resolution remote sensing image scene classification based on deep learning is summarized.(2)High-resolution remote sensing image scene classification with multi-level depth features.Traditional classification methods usually use a single bottom level feature or multiple feature combinations.However,the deep learning method generally only extracts image high level features.In response to this problem,this paper proposes a method of combining traditional bottom-level features with high-level features and then inputting neural networks for scene classification.Comparison of two scene classification methods,namely remote sensing scene classification based on multi-feature deep learning and remote sensing scene classification based on convolutional neural network.The experimental results show that the former can more effectively utilize the multi-level features of the image,thus achieving more accurate remote sensing scene classification.(3)High-resolution remote sensing image scene classification that incorporates multi-layer network features.The biggest advantage of the full convolutional neural network model U-Net is that the low-dimensional feature tensor of the underlying convolution is merged with the high-dimensional feature tensor extracted by thehigh-level convolution,and then input into the next convolutional layer.This paper proposes an improved multi-feature FCN model based on U-Net network.Compared with the original U-Net network,the fusion degree of low-dimensional and high-dimensional feature tensors is added.The BN algorithm is added to normalize the feature tensor,thereby reducing the amount of network computation.Using the Dropout operation,the network parameter redundancy is further reduced,which prevents the over-fitting phenomenon of the model training.The above improvements improve the accuracy of scene classification.