Deep Learning Algorithm For Remote Sensing Image Analysis And Big Data Platform Integration

With the rapid development of remote sensing observation platforms and new sensor technologies,remote sensing data has entered a significant data era.However,traditional remote sensing image processing methods can hardly handle such vast data.Remote sensing big data platforms have brought new opportunities for remote sensing data analysis and significantly improved the efficiency of large-scale remote sensing data analysis.Although these platforms have efficient management methods for multi-source and heterogeneous remote sensing data,they still lack efficient,intelligent analysis algorithms.Moreover,it is difficult to effectively integrate different algorithms due to the differences in development languages and environments,especially for the fast-developing deep learning algorithms in recent years,which lack the ability of plug-and-play and efficient integration.In response to the above problems,this thesis takes the extraction of buildings and photovoltaic panels in remote sensing images based on deep learning algorithms as an example.Furthermore,it conducts an in-depth study,proposes an optimization method for the performance of deep learning models,and investigates the encapsulation method of deep learning algorithms based on Docker containers and the integration method of big data platforms.The main innovation points of this thesis are as follows:(1)An improved Mask R-CNN instance segmentation model is proposed for the building extraction algorithm.Based on the residual neural network fusion convolutional attention model,a residual convolutional attention network is constructed to improve the feature extraction insufficiency problem.The loss function is optimized by adding the Dice Loss method,and then the feature learning process is optimized.And a post-processing strategy combining the Douglas-Peucker algorithm and Fine polygon regularization algorithm is introduced to make the building contours more regular.The experimental results show the advantages of this method.(2)The Filter-Embedded Network(FEPVNet)model is proposed for the PV extraction algorithm.This model embeds high-pass and low-pass filters and polarization selfattentiveness(PSA)into the high-resolution network(HRNet)to improve its anti-noise and adaptive feature extraction capability and finally enhance the accuracy of PV extraction.Meanwhile,three data migration strategies are introduced to migrate the FEPVNet model trained on Sentinel-2 images to Gaofen-2 images,thus improving the model’s generalization performance on a single data source for PV extraction in images at different scales.Finally,Validation of the effectiveness of cross-scale PV extraction through a series of ablation and comparison experiments.(3)To address the lack of intelligent analysis algorithms in remote sensing platforms,this thesis uses containerization technology to package and integrate the improved deep learning algorithms into the remote sensing big data platform,which improves the platform’s ability to support intelligent analysis.The large-scale PV mapping based on remote sensing big data platform is realized through the following four critical technology types of research:automated encapsulation of images,image management,container clustering,and container distributed computing.The extraction accuracy of targets in remote sensing images is improved by improving the deep learning model.Furthermore,based on integrating the deep learning algorithm in the remote sensing big data platform,large-scale detection target mapping is realized,thus improving the intelligence of the remote sensing big data platform.This study will enhance the intelligent understanding of remote sensing big data and play an essential role in extracting remote sensing information at global and regional scales.