Multi-level Classification Of Urban Land Cover And Structure Type From Landsat Remotely Sensed Imagery

Since the economic reform and open up,China's rapid urbanization process has greatly changed the spatial distribution of land cover and structure type,and has further influenced the structure of various terrestrial ecosystems,resulting in and aggravating a series of environmental problems,thus posing challenges to social sustainable development.Therefore,it is of great research value to improve the capability for timely and reliable classification of urban land cover and structure type.Landsat remote sensing imagery has become one of the most widely-used earth observation data sources,ascribed to its global coverage,free access,fine spatial resolution,short revisiting period and mature data support.Urban land cover includes artificial built-up areas and natural surfaces,and the information provided by land cover types is limited.The type of urban structure is mainly formed by the horizontal and vertical distributions of built-up areas,which also show significant influence on the rational urban planning.Therefore,this paper has carried out the multi-level classification research of urban land cover and structure type from Landsat remotely sensed imagery.Aiming at a series of problems in the classification of urban land cover and structure type,such as limited information provided by spectral bands,interference of unclear observations,and lack of height information,this paper develops a multi-level image classification framework composed of three components: the built-up area,the natural land cover,and the built-up structure,according to the characteristics of Landsat.The main contents of this paper are summarized as follows:(1)The built-up area(BUA)extraction based on Landsat is interfered by the bare soil,and the nighttime light(NTL)image is a unique and objective reflection of human activities.Therefore,their joint use has the potential to enhance the BUA extraction.Compared to the previous generation NTL images,i.e.,NPP/VIIRS(NPP),the spatial resolution of Luojia 1-01(LJ1)is more consistent with Landsat.However,the researches concerning the ability of LJ1 on BUA extraction is scare.Therefore,this study aims to achieve the quantitative evaluation of LJ1(and the combination of LJ1 and Landsat)for BUA mapping over the Mainland of China,across multiple biomes and city levels.The results show that the user's accuracy of LJ1 is generally higher than that of NPP,due to the higher resolution of LJ1.On the other hand,NPP gives higher producer's accuracy than LJ1,since the coarser resolution and blooming effect of NPP can actually lower the omission.Overall,LJ1 data can achieve more accurate mapping results in developed areas,such as megacities and large cities.When combing Landsat and the two types of NTL data,not only the complementary features of these two kinds of NTL data are captured,but also the daytime and nighttime information of BUA.Therefore,this paper combined Landsat,LJ1 and NPP data to generate the BUA datasets over the Mainland of China.(2)There are various types of land cover with obvious temporal features in natural surfaces.The imaging methods of Landsat and satellite faults cause the observations to be disturbed by clouds,cloud shadows,snow and ice,and SLC-off.These unreliable observations hinder the accurate representation and recognition of temporal features.In this paper,we present a novel co-training classification approach consisting of two steps to cope with the unclear observations.Firstly,we develop a method called the MCCR classifier where the unclear observations in the training samples are recovered by the use of the matrix completion(MC)algorithm,and the collaborative representation(CR)classifier is exploited to handle unclear ones in the unlabeled data.Secondly,considering that the random forest(RF)classifier is able to cope with contaminated data in an error tolerant way,a co-training approach(CotrRM)based on RF and MCCR is also proposed to further improve the classification efficacy.The CotrRM method is executed by iteratively constructing semi-labeled training sets based on the crisp and soft predictions of the two individual classifiers on the unlabeled data.The experiments on Wuhan city showed that MCCR performs as well as RF for the mapping of urban land cover with contaminated LTS imagery.Moreover,the CotrRM method has the ability to further improve the classification performance.The proposed approach can not only work effectively in the classification,but can also recover the unclear observations in the LTS imagery.(3)The structure of urban built-up areas includes the density and height distribution of buildings,and the ability of Landsat in the supply of building height information is limited.However,the echo intensity of SAR radar image is closely related to the surface geometry,and hence the integration of Landsat and SAR is employed in this paper to identify the built-up structure classification(BUSC).Considering the complexity of BUSC and the ability of deep learning to excavate high-level features,firstly,3D-CNN and Dense Net models are respectively designed for Landsat time-series images and Gaofen-3 SAR images,to extract their respective deep features.Then the depthwise separable convolution network is employed to effectively fuse these two deep features,and further learn the classification information of built-up structures.Experiments in Beijing and Shanghai show the injection of Gaofen-3 SAR can improve the classification accuracy of built-up structures,especially of the building height types.The proposed two-branch deep fusion method(FuseNet)has the ability to deeply integrate Landsat and SAR images to achieve reliable classification of built-up structure.