Research On Building Extraction From High Resolution Remote Sensing Image Based On Deep Learning

With the advancement of earth observation technology and remote sensing imaging techniques,a large amount of high-resolution remote sensing images can now be obtained.It is urgent to improve the intelligent utilization of remote sensing images for earth observation tasks.In recent years,with the popularization of high-performance graphics cards and the development of deep learning,deep learning technology has rapidly taken the lead in image recognition and image segmentation.In this context,this thesis explores the role of deep learning in intelligent earth observation using high-resolution remote sensing images as the data source,deep learning as the method,and buildings as the target for extraction.In view of the coexistence of large and small buildings in the current task of remote sensing image segmentation,Deep Lab v3+ uses atrous spatial pyramid pooling to solve this problem.However,in experiments,it was found that when a remote sensing image contains both large and small buildings,the segmentation of large buildings may have missing edges and incomplete segmentation.Therefore,attention mechanism is introduced to obtain the feature values in the context.This thesis compares the use of 9 attention mechanisms and parallelly uses them with the atrous spatial pyramid pooling layer in the Deep Lab v3+ network.On the WHU and Massachusctts public datasets,the network with added attention mechanism has a maximum increase in Io U of 3.85% and a minimum increase of 1.80%.Visually,it can be seen that the network without attention mechanism has poor segmentation effect on large buildings when segmenting images containing both large and small buildings,while the network with improved channel attention mechanism and dual-channel attention mechanism has more complete segmentation of large buildings.The experiment shows that adding attention can improve the segmentation performance of the network.When training models to extract buildings from remote sensing images,a large number of annotated images are needed for early training work.Although annotated remote sensing data can be obtained from open sources,the resolution of the remote sensing images varies,and building styles and layouts differ between different cities.Therefore,manual labeling work is still required when extracting buildings from specific areas,posing a challenge of insufficient labeled samples in some remote sensing fields.It is necessary to study deep learning methods under small sample sizes.This article uses the improved Deep Lab v3+ network to extract buildings from small sample remote sensing images by applying transfer learning.The pre-trained model on public datasets is transferred to small sample datasets for frozen and unfrozen training,adjusting the model using limited labeled samples to obtain higher semantic segmentation accuracy.