Object Detection And Change Detection Of Remote Sensing Images Based On Convolutional Neural Network

Automatic detection of strategic objects,such as harbors and oil depots,and land use change detection based on remote sensing images have always been important aspects of Remote Sensing applicable research.However,with the resolution enhancing of remote sensing images,the existing object detection and change detection algorithms are facing great challenges.The reason is mainly that the existing detection algorithms mostly depend on handcrafted features,which results in the poor robustness of these algorithms.In recent years,CNN(Convolutional Neural Network)with great feature extraction ability has achieved great success in the image classification application.How to use CNN to extract features suitable for objectdetection and change detection to improve the accuracy of objectdetection and directly detect land use change and change types has great research and application value.Therefore,on the basis of analyzing and discussing commonly used CNNs and siamese networks,the master thesis focuses on the detection of harbors and oil depots and change detection in large scale optical remote sensing images based on convolutional neural network.The main work and contributions of this thesis can be summarized as follows:(1)For harbor detection in a large-scale remote sensing image,the prior knowledge of harbors always locating at the boundary between land and sea,and consisting mainly of wharves is utilized.A harbor detection method combing the CNNs and structural features is proposed.At first,the Residual Network(ResNet)is applied to detect the coastal zones in downsampled low resolution remote sensing images to identify the range of areas where harbors may exist and reduce false alarms in the follow-up detection.Then the Single Shot MultiBox Detector(SSD)is adopted to detect the wharves.A semi-closed structure detector is designed to verify those true wharves according to their geometric structure,which further reduces misidentification.Finally,the detected areas are combined and the harbors are found.Compared with the harbor detection methods of directly detecting wharves,the proposed method improves the precision and recall rates by 10.25% and 2% on the average,respectively.(2)For oil depots detection in a large-scale remote sensing image,an oil depots detection method fusing multi-level features is proposed.The ResNet is applied to detect the regions containing oil depots in downsampled low resolution remote sensing images firstly.Then multi-level features are fused to detect and locate oil depots precisely.An improved DBSCAN cluster algorithm is used to obtain different oil depots.The proposed method has 2.2% and 0.9% higher than SSD method in terms of recall and F1-measure,respectively.(3)For change detection of remote sensing images,a supervised convolutional neural network architecture named AggregationNet is proposed,in which the change detection task is transformed into a classification task,however,different from the change detection after classification,it can not only detect land use change,but also dierctly obtain change types.AggregationNet consists of a feature extraction network and a feature aggregation network.And a two-dimensional label map is designed to train AggregationNet.AggregationNet achieves 1.6% higher than post-classification method in term of accuracy on the SAT-6 dataset.Compared with the traditional CXM(Conditional multilayer mixed MRF model),SCCN(Symmetric Convolutional Coupling Network)and DSCN(Deep Siamese Convolutional Network)based on CNN on the TISZADOB dataset,AggregationNet obtains the highest precision and recall rate.