A Research On Remote Sensing Intelligent Interpretation Of Urban Functional Zones Via Multimodal Deep Learning Models

Urban functional zone refers to urban spaces divided according to the distribution of various human activities and social services.It is also the basic unit of urban planning and resource allocation.Accurate and up-to-date spatial distribution of urban functional zone can support urban policy formulation and planning.Urban functional zone interpretation aims to map urban spaces to different functional zone categories.It is an important task for urban management and sustainable development.There are difficulties in the intelligent interpretation of urban functional zone: few labeled samples,strong heterogeneity,and highly mixed regions.In urban areas,especially mega cities,the area of the city is wide,resulting in a smaller sample size can be obtained.The existing methods for extracting functional zones based on fully supervised method can not accurately extract urban functional zones under a few labeled samples.Besides,urban functional zones and their subcategories have inconsistent spatial characteristics at different spatial locations,showing strong spatial heterogeneity.In addition,as a result of historical planning(such as policies,renovation of old urban areas,etc.)and geographical factors(water resources,topography,etc.),urban scenes are fragmented and have the characteristics of "highly mixed regions",that is,in the same area,usually includes a variety of functional zones categories.The characteristics of "highly mixed regions" in functional areas make it difficult to further fine-grained semantic segmentation of the location and boundaries of urban functional areas and their subcategories in mixed functional zones.As a solution,combining artificial intelligence algorithms represented by deep learning with remote sensing big data and geospatial big data is an effective way to improve the interpretation of urban functional zones.To address the above issues,the main research contents of the thesis are summarized as follows:(1)A semi-supervised learning strategy-based interpretation method for urban functional zone classification: To address the difficulty of "few labeled samples," a semisupervised learning strategy that automatically selects pseudo samples from a large number of unlabeled samples is proposed.Considering the strong spatial heterogeneity of urban functional zones,it is difficult to fully capture the characteristics of functional zones using only remote sensing or social remote sensing data.This thesis proposes a bibranch neural network model to improve the accuracy of urban functional zones interpretation by comprehensively utilizing the different characteristics of remote sensing images and social sensing data.(2)Image classification methods for informal settlements mapping by fusing remote sensing images,time-series population density data,and street view images: Aiming at the difficulty of "strong heterogeneity," this thesis takes the subcategory of residential areas(informal settlements)in functional areas as an example and combines multiple modal data sources and multimodal deep learning methods.An urban informal settlement classification model based on remote sensing images and time-series population density data and an urban informal settlement classification model that fuses remote sensing images and street view images are proposed.(3)Multimodal semantic segmentation methods for subcategories of urban functional zones based on multimodal deep learning models: To address the difficulty of "highly mixed regions," this thesis takes the subcategory of residential areas(informal settlements)and the subcategory of public service areas(urban open space)in urban functional zones as examples and combines remote sensing images,building polygon data and multi-spectral remote sensing images.A multimodal deep learning-based semantic segmentation method for urban informal settlements mapping and a multimodal deep learning-based semantic segmentation method for urban open spaces are proposed.In summary,this thesis employs multimodal deep learning methods as the main methods and remote sensing big data,and geospatial big data as the main data sources to interpret urban functional zones and their subcategories.This thesis can provide new insights and novel methods for the interpretation of urban functional zones and their subcategories,and has great scientific and social significance.