Research Of Object And Land Cover Extraction Based On Combination Of Visible And Infrared Remote Sensing Images

Object and land cover extraction is an important task in the interpretation of optical remote sensing images and plays a huge role in the civil and military fields.With the continuous development of space-based optical remote sensing technology,the optical imaging spectrum has been continuously expanded from visible light to infrared,and visible light and infrared composite imaging technology has become an important development direction of space-based optical remote sensing,which greatly facilitates the simultaneous acquisition of visible and infrared remote sensing images.Visible and infrared remote sensing images have strong complementarity,and the joint application of the two types of images for object and land cover extraction has become an important research direction in remote sensing image interpretation.For object extraction,ship is a type of object that is focused on,and the existing deep learning ship detection algorithms are mainly based on visible remote sensing images for research,and a large number of false detections and missed detections will occur in complex scenarios such as cloud and fog interference.For land cover extraction,buildings,roads,water bodies and vegetation are the key types,and the existing deep learning land cover classification algorithms based on visible remote sensing images are not accurate due to the small spatial appearance differences between different types of land cover features.In order to solve the above problems,this paper proposes VI-YOLO ship detection algorithm and VIYOLO-seg land cover classification algorithm based on the combination of visible and infrared remote sensing images.The average precision of VI-YOLO algorithm reaches0.976 on the self-built ship dataset of visible and infrared remote sensing images,which effectively alleviates the false detections and missed detections in complex scenes,and the mean intersection over union of VI-YOLO-seg algorithm reaches 0.612 on the selfbuilt land cover classification dataset of visible and infrared remote sensing images,which outperforms the classical land cover classification algorithms.And the complementarity of visible and infrared remote sensing images and the effectiveness of the algorithms are verified.The main research work of this paper can be summarized as follows:Firstly,in order to solve the lack of ship dataset and land cover classification dataset combined with visible and infrared remote sensing images in the public datasets,the ship dataset and land cover classification dataset based on the combination of visible and infrared remote sensing images are constructed.Extract channels of the optical remote sensing data of Sentinel-2 satellite,slice and label the images to make a ship dataset for the training and testing of ship detection algorithm.Screen and extract channel of the public multispectral remote sensing image land cover classification dataset to make a land cover classification dataset for the training and testing of the land cover classification algorithm.Secondly,aiming at the problem that the performance of the existing ship detection algorithms based on visible remote sensing images degrades under complex background,a ship detection algorithm VI-YOLO is designed.Firstly,a dual-channel input network of visible and infrared images is constructed at the network input end to realize the combination of visible light and infrared remote sensing images;then,CSPDarknet-53 is used as the backbone network to extract spatial features and spectral features;finally,a fast spatial pyramid pooling module and SIo U loss function are introduced to speed up network convergence and improve network accuracy.Experiments show that the method effectively alleviates false detections and missed detections in complex scenes,and its performance is better than classical object detection algorithms.Thirdly,aiming at the problem that the existing land cover classification algorithms based on visible remote sensing images are not sensitive to the difference of land cover features,resulting in insufficient algorithm accuracy,the VI-YOLO-seg land cover classification algorithm is designed.Firstly,the pixel-level fusion of visible and infrared remote sensing images is realized by using the dual-channel input of visible and infrared images at the input end;then,a multi-scale feature fusion extraction network is used to extract land cover features of different scales at the same time to alleviate the decrease in algorithm accuracy caused by sample scale differences;finally,Proto Net is introduced as the segmentation head of the network to alleviate the algorithm performance degradation caused by class imbalance.Experiments show that the method improves the accuracy of land cover classification and outperforms the classical land cover classification algorithms.