Reasearch On Road Extraction Of Remote Sensing Images Based On Label Noise Dataset

The high-resolution remote sensing images provide abundant spatial and contex-tual information of ground object,which can reflect the subtle differences between dif-ferent ground objects accurately.This offers a great opportunity to study the character-istics of ground object.Road extraction has received much attention from researchers,because it is a fundamental task in remote sensing image processing and has a wide range of applications,including urban planning,disaster mitigation estimation,field rescue and military deployment.With the development of deep learning,deep net-works have achieved promising accuracy and efficiency in feature extraction and clas-sification.Therefore,the deep networks becomes major methods quickly in the field of remote sensing image road extraction.In existing research,the road extraction methods based on deep networks are mainly fully supervised and rely on numerous samples to achieve promise performance.However,obtaining a vast number of training samples is difficult since labeling remote sensing image is labor-intensive and time-consuming.As a result,the insufficient samples limit the application of deep networks in road ex-traction from remote sensing images.In recent years,to address the above issue,the existing crowdsourced vector data is utilize to label remote sensing images and build large-scale data sets.This method has effectively alleviated the"sample starvation"problem of deep road extraction models.However,large-scale data sets produced by crowdsourced vector data could contain plenty of label noise,i.e.,there are samples with inconsistent labels and actual roads,which would reduce the generalization ability of deep road extraction models.Label noise in road extraction is highly random and thus the existing methods of handling label noise is difficult to obtain satisfying performance.To this end,this paper conduct research to solve label noise issue by exploiting the spectral feature of remote sensing images,label posterior and multiple source data.And three methods are proposed to improve the robustness of deep networks for road extraction from remote sensing images.The contributions of this paper are summarized as follows.(1)A label noise tolerance method was proposed based on spectral similarity rep-resentation:Firstly,a spectral embedded label noise model was developed by explic-itly introducing ground object information contained in the spectral prior knowledge,which aims to improve the accuracy of modeling the label noise distribution.Then,a deep robust road extraction framework was proposed by combining the powerful fea-ture extraction capability of deep networks and label noise model.The label noise dis-tribution information is used in framework to reduce the contribution of samples to the optimization of model parameters,so as to enhance the tolerance of the deep road ex-traction model to label noise.Finally,a label noise regularization term was introduced to improve the optimization efficiency of the deep robust road extraction framework.Extensive experiments on three road extraction datasets showed that the method can effectively mitigate the effect of label noise and obtain better road extraction results.(2)A label noise learning method was proposed based on sequence posterior in-formation:This method made full use of multi-stage label posterior probabilities to improve the learning ability of deep road extraction model for label noise samples.Specifically,firstly,label probability sequence was obtained by collecting label poste-rior information from deep road extraction models with different optimization degrees.Then,an adaptive label correction method was proposed to identify mislabeled samples using the consistency of the label probability sequence,and correct the mislabel by min-imizing the distance between the label probability sequence and the possible labels.In addition,a sequence uncertainty regularization term was introduced to increase abil-ity of deep road extraction models to corrected samples.Experiments showed that the proposed method achieves to 11.86%improvement in F1 evaluation metrics compared with the label noise learning methods based on single-model.(3)A label noise refinement method was proposed based on multi-source vector data:Since the multi-source vector data,such as OSM,ZMap and GPS was often inde-pendence and complementarity,the proposed method combined them to re-label high-resolution remote sensing images,which aims to suppress label noise and optimize sample quality.Meanwhile,Multi-Map Integration Model(MMIM)was proposed to exploit the real label sample distribution in the multi-source vector data and generate high-precision refined labels.Finally,the refined labels were used to train deep road ex-traction model and reduce the negative effective of label noise.The experiments were performed on the main urban area of Zhengzhou city,which covers about 1000km~2.The proposed method can extract the road regions form study area in 8 hours and the F1 evaluation metric can achieve 78%.These experimental results demonstrated the potential advantages of the proposed method in practical applications.