High-resolution Remote Sensing Image Water Extraction Based On Multi-scale Feature Enhancement Network

It is becoming more and more important for water resources investigation,water conservancy planning,flood control,disaster reduction,ecological protection,and sustainable development to study the water-body spatial-temporal distribution and obtain water-body information quickly,accurately,and efficiently.With the progress of remotely sensed technology,the characteristics of instantaneous imaging,rich spectral information,and wide coverage range of remotely sensed.It can obtain the distribution information of different ground objects in large spatial-temporal scale.Remotely sensed technology has become the primary means of water-body information extraction.Especially in recent years,with the establishment of earth observation systems with different purposes,spectra,and resolutions,remotely sensed data has exploded.It provides rich data sources for water-body information extraction research.High resolution remotely sensed imagery has attracted lots of scholars' attention to water-body information extraction due to its clear geometric structure,rich texture information,diverse spectral features,and high temporal resolution.Traditional water-body extraction methods are mainly based on the spectral difference of ground objects,using prior information or intensity threshold to distinguish water-body from non-water-body.Based on medium and low resolution remotely sensed imagery,accurate extraction of water-body can be achieved.Due to the loss of spatial context information,water-body extraction results are broken and contain a lot of "salt and pepper" noise when processing high resolution remotely sensed imagery.At the same time,these kinds of algorithm are highly dependent on the scene,complicated processing process,and low degree of automation.It is difficult to meet the requirements of processing quantitative and diversified remotely sensed imagery big data.With the rapid development of deep learning theory,the deep convolutional neural network algorithm is proposed,which automatically learn multi-level and multi-scale features by supervising the training process.Based on deep convolutional neural network,water-body extraction from high resolution remotely sensed imagery is realized.By constructing a nonlinear model between imagery and water-body features,the semantic characteristics of water-body are expressed end-to-end,so as to realize water-body information extraction.However,how to construct one deep convolutional neural network,which not only fully consider the context information of different levels-scales,but also enhance the water-body features to improve the water-body extraction accuracy.The network has strong feature expression ability.At the same time,the deep convolutional neural network with high efficiency and strong training ability has become an important research topic.Therefore,this study proposes a multi-scale feature enhancement network for water-body extraction from high resolution remotely sensed imagery.The method mainly includes data preprocessing,water-body data set construction,and multi-scale enhancement network model design.The main contributions lie on:(1)Water-body data set construction of high-resolution remotely sensed imagery.The data processing of high-resolution remotely sensed imagery,such as data preprocessing,imagery label,data set division,etc.,are studied,GOGF and Love DA water-body data sets are constructed.It provides data set support for water-body extraction research of high-resolution remotely sensed imagery.(2)An end-to-end water-body extraction algorithm based on multi-scale feature enhancement network is proposed.For convolutional neural network,poor feature expression ability,weak training ability,and other problems.Multi-scale feature extraction module Res2 Net and up-sampling feature decoding module are designed to construct the encoderdecoder network.At the same time,the strip pooling feature fuse module and context feature extractor module are integrated into the network to make full use of context information at different scales and levels to enhance the expression of water-body semantic information.(3)The performance of multi-scale feature enhancement network for water-body extraction from high-resolution remotely sensed imagery is conducted.Based on GOGF and Love DA data sets,the influence of different hyperparameters on water-body extraction accuracy is discussed.The advantages of the proposed method for water-body extraction from high-resolution remotely sensed imagery are proved by ablation and comparison experiment.