Research On Urban Impervious Surface Mapping From Multi-source Remote Sensing Images Based On Multiple Kernel Learning Model

With the rapid urbanization,urban development is facing tremendous ecological pressure.Impervious surface refers to artificial features which impede the infiltration of water into the soil.It has been widely recognized as an indicator of urbanization level and environmental quality.Therefore,the dynamic monitoring of impervious surface distribution has become urgent for assessing the impact of urbanization on urban ecological environment.Remote sensing,with its convenience and repeat coverage over large geographic areas at low cost,has been intensively investigated for impervious surface mapping.Optical remote sensing images are widely used in impervious surface estimation.However,the urban landscape is highly spatial heterogeneous,the urban land cover is complex,and the impervious surface materials are diverse.The problem of mixed pixels in medium and coarse resolution imagery,spectral confusion between different land cover and the influence of cloud cover make accurate impervious surface extraction still facing challenges.Research shows the use of multisource remote sensing images,especially SAR and nighttime light(NTL)data,can provide complementary information for accurate impervious surface mapping.In this paper,the synergistic use of medium-resolution Landsat-8 optical image and Sentinel-1 SAR data is suggested for urban impervious surface mapping at the sub-pixel level in Anyang city,in addition,the synergistic use of Landsat-8 and Luojia-1-01 nightlight data is suggested for large-scale impervious surface mapping in Jiaozhou Bay.A novel multiple kernel learning(MKL)framework is proposed to integrate heterogeneous features from optical and SAR data more effectively.And the impervious surface abundance of the study area is estimated by applying the developed multiple kernel support vector regression(MKSVR)model.The experimental results and conclusions are as follows:1.Different feature extraction methods are emplyed to fully mine the potential information in optical and SAR data,and the random forest(RF)algorithm is used for feature selection.The features extracted from optical and SAR images include:Landsat-8 spectral information and gray level co-occurrence matrix(GLCM)texture features;Sentinel-1 VH/VV backscatter intensity,H-?target decomposition parameters and GLCM texture features.2.Features extracted from optical and SAR have different statistical characteristics,dimensions and physical meaning.Conventional layer-stacking approach can not make full use of the rich information with different characteristics.Therefore,we propose an effective MKL framework to integrate optical and SAR data.The proposed MKL framework with differential evolution(DE)based parameter selection,is a weighted linear combination of multiple basis kernels corresponding to each group of features.It is expected to fuse heterogeneous features more effectively and provide improved performance.3.Considering the mixed pixels problem in Landsat-8 imagery,impervious surface abundance of the study area are estimated by employing the established multiple kernel support vector regression(MKSVR)model.The impervious surface ground truth at a sub-pixel level is derived from a high resolution image by means of object-oriented classification.Compared with the single kernel SVR(SKSVR)and composite kernel SVR(CKSVR)model,the MKSVR model achieved the highest accuracy of impervious surface mapping,the root mean square error(RMSE)is 0.2031 and the coefficient of determination(R~2)is 0.8321.The results show that the incorporation of optical and SAR does not guarantee the improved performance,and the construction manner of kernel function is important.In addition,MKSVR model achieves a noteworthy improvement for impervious surface estimation compared to that using optical image alone,the RMSE is decreased by 4.30%,and the R~2 is increased by 9.47%.The synergistic use of optical and SAR data reduces the spectral confusion between bright impervious surface and bare soil,as well as between dark impervious surface and shadow,water.4.Combining Landsat-8 and Luojia 1-01 images,three different methods are used for large-scale impervious surface mapping around Jiaozhou Bay,namely simple threshold segmentation method,light index method(HSI and MNDISI index)and SVM supervised classification method.Among them,the threshold method can not reflect the spatial structure information of urban area very well,and is affected by the lack of light and noise.HSI and NDISI index can reflect more abundant spatial details of urban area.By combining Landsat data,the extraction accuracy of impervious surface is greatly improved.However,confusion between beach area and impervious surface is still serious.The integrated Luojia1-01,NDVI index,and land surface temperature(LST)SVM classification method(INNL-SVM)achieves the highest accuracy among the three methods.Compared with that Landsat-derived impervious surface result,the overall accuracy(OA)is increased from 94.57%to 96.04%and Kappa value is increased from 0.8914 to 0.9207.INNL-SVM uses threshold method for automatic sample collection,which solves the shortcomings of the traditional method,for the size of sample data is limited and the collection is difficult.The SVM method with automatic sample collection improves the accuracy and efficiency of impervious surface extraction.The experimental results show that,compared with VIIRS-DNB,Luojia 1-01 night light data can provide more detailed information of urban spatial structure and has great potential for large-area impervious surface mapping.The synergistic use of Landsat-8 and Luojia1-01 is an effective method for regional impervious surface mapping.