Research On Classification Of UVA Remote Sensing Image Based On Semantic Segmentation Model

With the continuous development of computer vision technology,semantic segmentation technology has been widely used in many fields,such as driverless,medical image analysis,remote sensing image processing and so on.At present,the acquisition of remote sensing data mainly adopts the traditional satellite remote sensing technology and UAV remote sensing technology.The traditional satellite remote sensing technology covers a wide range,but it lacks real-time,long revisit cycle,is vulnerable to the influence of clouds,and it is difficult to obtain high-quality remote sensing images;The UAV remote sensing technology has advantages in image resolution and timeliness,but most of them need to be obtained manually.Therefore,taking the current situation of campus land use as an example,this paper uses the UAV remote sensing image as the basic data,and uses the improved semantic segmentation model to achieve good results in UAV remote sensing image classification.The main contents are as follows:(1)Multispectral UAV is used to take aerial photos of the campus in summer and winter to obtain UAV remote sensing images.Set the target category of semantic segmentation as5 categories,and the label samples are made by Arc GIS software.At the same time,in order to ensure the full training of model parameters and enhance the generalization ability of the model,the UAV images and corresponding label samples are expanded by random cutting,adding point noise and geometric transformation.(2)Based on the self-made UAV remote sensing image data set,four models of SVM,Seg Net,Deeplabv3 and Deeplabv3+ are used for experimental comparison.The results show that the prediction accuracy of the three classical semantic segmentation models is better than that of the SVM.Among them,Deeplabv3+ semantic segmentation model has the best effect,and the CPA reaches 87.94%.(3)Aiming at the problems of unclear edge segmentation and breakpoints in deeplabv3 +model,the model is improved: by embedding convolution attention mechanism(CBAM)in the backbone network,the influence of irrelevant features on accuracy is reduced;(2)by introducing deep separable convolution into ASPP module and adding a convolution branch,the amount of module parameters is reduced and the training efficiency and performance of the model are improved;(3)by adding residual attention module at the Decoder,It can effectively improve the problems of accuracy reduction and gradient disappearance caused by too deep network layers,and make the model convergence faster and more stable.In order to verify the recognition effect of the improved model,taking the main campus of Anhui Agricultural University as the experimental area,two groups of comparative experiments in summer and winter are designed.The results show that the PA and CPA of UAV image in summer are increased by 7.28% and 6.84% respectively compared with the original deeplabv3+ model;Compared with the original deeplabv3+model,the PA and CPA of UAV images in winter are increased by 11.44% and 11.24%respectively;In winter,the image recognition accuracy of roads and buildings is higher.(4)The improved Deeplabv3+ model is used to classify the campus.In order to more accurately obtain the distribution of evergreen plants and deciduous plants,the NDVI map of multispectral UAV is combined with the identification results.It is obtained that: the old campus in phase I is dominated by buildings and plants,the new campus in phase II is dominated by other types,mainly playgrounds and activity areas,and the planned area in phase III is dominated by grassland,mainly experimental fields.