Scene Classification And Semantic Segmentation Of Optical Remote Sensing Image Based On Deep Learning

With the improvement of earth observation technology,the number of optical remote sensing images has increased sharply.A vast number of optical remote sensing images provide essential data for scientific research and social development.However,the processing of many optical remote sensing images consumes huge material,human and financial resources.Therefore,it is urgent to solve how to extract useful information from optical remote sensing images accurately,efficiently,and intelligently.In recent years,it has promoted the development of optical remote sensing image analysis tasks,including scene classification and semantic segmentation,that deep learning and its related technologies have been successfully applied in remote sensing.This paper focuses on applying scene classification and semantic segmentation based on deep learning in optical remote sensing images.For a large number of optical remote sensing images,the scene classification model is devoted to screen the target images that need attention.The detailed information of the region of interest in the target image is extracted through the semantic segmentation model,to realize the intelligent extraction of practical data in optical remote sensing images.1.Aiming at the problems that the scale of scene classification model is too large and some confusing scenes are difficult to classify accurately,which is not conducive to terminals deployment,so a scene classification model based on lightweight network and integrated learning is constructed.The model optimizes the structure through lightweight network,compresses the scale of the model,and reduces the number of parameters of the model.The model uses ensemble learning to construct multiple classifiers,and extracts features through multiple classifiers to enhance the classification performance of the model.The experimental results show that this model only trains 20% of the samples on AID and NWPU-RESISC45 datasets,and achieves the accuracy of 94.32% and 93.36%.As the confusing scene,the accuracy of school and commercial areas is more than 4% higher than that of the suboptimal model,and the number of parameters and floating-point operations is low.Compared with the classical scene classification model,this model not only improves the accuracy,but also reduces the requirements of computing environment configuration,and can classify confusing scenes more accurately.2.Aiming at the problem that the semantic segmentation model of optical remote sensing image based on deep learning cannot accurately segment small-sized objects and object boundaries,a semantic segmentation model of optical remote sensing image combining attention and U-shaped network is constructed.The model extracts and fuses multi-level features through densely connected structures,which effectively retains the feature information of small-sized objects;The attention mechanism is used to improve the ability of the model to analyze the context information,so that the model pays more attention to small-sized objects and object boundaries.The experimental results show that the Io U and F1 scores of the car as the small-size object in the Vaihingen data set are 0.719 and 0.901,which are 16% and 11% higher than those of the suboptimal model.Compared with the standard semantic segmentation models,the object boundaries in the segmentation results of this model are more precise and more accurate.Even if the object is disturbed by shadow,occlusion,or insufficient light,it can be accurately segmented.3.The Self-built optical remote sensing image dataset is produced.A system for scene classification and semantic segmentation of optical remote sensing images is realized.According to the requirements of the task,scene classification dataset of optical remote sensing image is built,which consists of three kinds of scenes: Town,port,and other.And the semantic segmentation dataset of optical remote sensing image is constructed which consists of three kinds of objects:water area,ship,and background.Self-built datasets and public datasets jointly verify the performance of the model.The scene classification and semantic segmentation system for optical remote sensing images is designed based on Python and Py Qt.The system integrates the scene classification model and semantic segmentation model constructed in this paper,and displays the results and evaluation indexes of the model through the visual interface.