Research On Impervious Surface Coverage And Change Information Mining Methods In Large-scale And Long Time Series

Since the 21 st century,the impervious surface area in China has grown rapidly with the develop of social economy.High precision and frequency recognition of impervious surface cover and changes is significant for research on ecological environment and social sciences.Moreover,data type and quantity and the data processing performance have been greatly improved in the current big data era,contributing to large-scale information mining with multi-source remote sensing data becoming a hot topic.Due to the low acquisition cost,high standardization,and complete spatial and temporal coverage,low-and medium-resolution data are widely used for large-scale research on both coverage and change of impervious surfaces or other land-surface.However,for the object of impervious surface,owing to the high intra-class complexity in medium-resolution images and small inter-class difference,it is not easy to interpret for experts even though abundant data and features are available.Thus,high-precision information mining is limited.This thesis concentrates on the impervious surface cover and changes in China and develops the information mining methods based on the dataset from Landsat satellite,Chinese Gaofen series satellite and others.Detailly,a high accuracy labeled dataset was built;a deep learning model for extracting impervious surface cover from multi-sourceinput was presented;an automatically framework for mining impervious changes from long time series data was designed;more reliable impervious surface cover and change maps in large-scale were generated by the zonal strategy.The main achievements and conclusions are as follows:(1)Developing impervious surface cover and impervious surface change dataset based on multi-source and multi-temporal data.Because of the inadequate medium-resolution images features,2m-resolution domestic high-resolution satellite products were introduced as a spatial supplement besides Landsat and terrain data.The total area of impervious surfaces is small,of which spatial distribution is also extremely uneven.Moreover,the previous surface type is complex.According to these characteristics,a stratified sampling method based on information entropy constraints was proposed,and a more representative sample points set was obtained by sampling in areas with high heterogeneity.Finally,a dataset of impervious surface cover containing about 200,000 samples was built by visual interpretation.Another one that contains change information was generated automatically with the strategy of constrained by a continuous change detection algorithm.Both datasets were built to help dig out the impervious surface changes without manual intervention.(2)A deep learning method for impervious surface cover information mining based on multi-feature fusion was proposed.Deep learning methods have powerful feature extraction capabilities.This study adopted the idea of feature-level fusion to increase the separability of impervious surfaces through advanced feature expression.Firstly,applied the Res Net model to the high-resolution dataset to express advanced spatial features;Next,applied the MLP model to the medium-resolution dataset to express advanced features based on spectral and topographic information;Finally,combined the features from different sources to dig out high-precision impervious surface cover information.In addition,the two-class classification problem was turned into a 4-class classification task,which can avoid the problem that the differences between classes are small;A fuzzy class strategy was also used to reduce the negative impact of ambiguous samples in the dataset.The results show that our model’s ability to extract impervious surfaces is superior to any other existing product,and the precision and recall were both up to 0.93.(3)Developed a spatiotemporal deep network model for impervious surface change information.A scheme for changing information mining based on classification was designed.In order to solve the problem of insufficient information utilization in information mining of long-time series data,a hybrid model was designed by combining CNN and LSTM,which can extract the spatial and spectral features of time-series images.Furthermore,a time window strategy was proposed to realize data augmentation and improve the model’s scalability;Besides,the neighbor adjusted cross-entropy loss was proposed to cope with the complex impervious surface growth mode and the noise in the data set.The result is tested on an independent validation set,and shows that the proposed model has higher noise immunity,which leads to higher validation overall accuracy up to 85%.(4)A multi-model large-scale impervious surface coverage and change mapping scheme was designed based on landscape context zoning.China has a vast territory,and the topography,culture,and climate of different regions are quite different,resulting in differences in data distribution and patterns of impervious surfaces.Especially for the central and western regions,the impervious surface is sparsely distributed and easily confused with the bare soil background,which is significantly different from other regions.In order to improve the extract accuracy of overall impervious surface cover and change information,a zonal method based on landscape context features was designed.It contains multiple trains models for different zones to achieve high-precision national impervious surface mapping.Constrained by the high-precision result,the false changes in the pervious surface area were removed.Then the final large-scale impervious surface change mapping results were obtained.Finally,the high accuracy impervious surface cover product of China in 2020 and the impervious surface change product of China from 2000 to 2020 are generated.