Super-resolution Surface Linear Water Body Mapping With Remotely Sensed Imagery Using Convolutional Neural Network

Surface water bodies play a crucial role in regional and global hydrological and biochemical cycles,accurate and rapid extraction of the spatial distribution of them is of great significance for the study of biodiversity and climate change.The development of remote sensing sensors with different spatial and temporal resolutions provides abundant data base for the acquisition of water bodies.In practice,due to the mutual restriction of spatial and temporal resolutions of remote sensing sensors,a single remote sensing sensor cannot meet the demand of rapid monitoring of dynamic change of water bodies with high spatial and temporal precisions.Super-resolution mapping(SRM)is a scale transformation method for remotely sensed imagery which can consider the land cover change,it converts a coarse-resolution remote sensing image into a corresponding fine-resolution land cover map through the processes of spectral unmixing and spatial downscaling of land covers.SRM can be an effective approach for the dynamic monitoring of water bodies at fine scale.Surface water bodies mainly present planar and linear shapes,compared with planar water bodies,linear water bodies possess more complex spatial characteristics.Current SRM methods can achieve good results for planar water bodies,but research on how to extract spectral and spatial features of linear water bodies is scarce.There are still several challenges in super-resolution linear water body mapping:1)Traditional SRM methods have difficulties in representing the spatial distribution characteristics of linear land covers;although deep learning-based model improved the accuracy of spatial feature representation,it still has difficulties in maintaining spatial continuity for linear features.2)Most spectral unmixing models are based on linear unmixing,which is hard to describe the spectral mixing mode of complex land covers fully,leading to large fraction errors.This would bring an adverse impact on the process of downscaling linear water bodies.This dissertation takes river water bodies and leads as research objects,studies the above problems using deep learning,which has powerful feature extraction ability.First,the issues concerning super-resolution river water body mapping,that is,how to extract spectral and spatial features,are analyzed and studied;then,the super-resolution lead mapping and the application of the corresponding SRM results in the monitoring of turbulent heat flux are explored.Specifically,the main research contents of this dissertation are as follows:(1)Due to different geographical environments,river water bodies present complex linear spatial distribution characteristics,which is hard to be fully described by traditional spatial feature representation models since they are often based on maximum spatial correlation hypothesis.This dissertation proposes a nested multi-layer feature fusion convolutional neural network(CNN)-based spatial feature description model for the river water bodies.The model takes river water fraction image as input,and automatically extracts the complex spatial distribution features of river water bodies through training.An adaptive cross-entropy loss function combining fraction information is proposed to improve the model’s ability to maintain small river water bodies’ connectivity.Experiments show that the proposed method can effectively extract the multi-level spatial features of river water bodies,which overcomes the limitations of existing methods in processing linear features to some extent(2)Traditional super-resolution water mapping methods takes spectral unmixing and spatial downscaling as two independent processes,which will lead to the propagation of river water fraction error to the mapping results and cause large uncertainty in the results.This dissertation proposes a cascaded spectral-spatial CNN model for super-resolution river water body mapping.The model takes the coarseresolution remotely sensed imagery as input directly,contains two sub-networks to extract the spectral and spatial features of river water bodies,reduces the influence of spectral unmixing error by constructing a fraction loss function,and optimizes the whole model’s parameters by combining the category loss objective function.Experiments show that the proposed method can consider the unmixing error fully,and improve the river water mapping results effectively.(3)The spatial distribution of river water body is highly correlated with topography,introducing topography information as a spatial constraint in the super-resolution river water body mapping model can improve the accuracy of the results.Since fineresolution water occurrence data can provide accurate topography information of river areas,this dissertation proposes a remotely sensed multi-spectral imagery and water occurrence data-combined super-resolution river water body mapping framework,the framework uses multi-spectral imagery to provide spectral and spatial prior information,and uses water occurrence data to constrain the spatial distribution of river water body.In addition,this dissertation explores how to fuse the water occurrence data to take full advantage of them.Experiments show that the combination of water occurrence data is important for improving the accuracy of river water mapping results.(4)Leads are another important surface linear water body form,precise monitoring of them is of great significance for the study of the arctic climate change.Different to traditional river water bodies,leads and surrounding sea ice areas have greater difference in surface temperature than in other spectral bands.Therefore,the ice surface temperature retrieved from thermal bands is an optimal data for leads identification.Since retrieving ice surface temperature from thermal bands is greatly affected by the mixed pixels of thermal bands,it is often hard to identify small leads.This dissertation proposes a super-resolution leads mapping model for producing fine-resolution leads from coarse-resolution remotely sensed data.The model takes ice surface temperature data as input,and uses a multi-layer feature extraction and fusion CNN to extract the surface temperature and spatial features of leads and form the nonlinear relationship between input and output data.In addition,the application potential of the SRM results in the estimation of subpixel turbulent heat flux over leads in the arctic region is studied.Experiments show that subpixel scale analysis method can effectively improve the accuracy the estimation accuracy of turbulent heat flux.