Research On Application Of Risk Sources Monitoring In Water Sources Using Deep Convolutional Neural Network

The extraction and quantitative evaluation of risk sources surrounding water source area is an important technical link for realizing water source area monitoring and management.The improvement of the spatial resolution of remote sensing images makes the surface objects presented with more detailed texture,clear shape and spatial distribution information.And the convolutional neural network can intelligently extract a variety of features on high-resolution images.The convolutional neural network is able to accurately locate risk sources surrounding water sources and extract them,and based on the extraction results,we can quantitatively evaluate the risk sources surrounding water source areas,which can help the relevant departments to discover hidden dangers in the water environment in time.This paper conducts research from three aspects: determining risk sources and obtaining and producing remote sensing images for the task of extracting risk sources,extracting risk sources of water source from high-resolution remote sensing images accurately,and quantitatively evaluating non-point risk sources surrounding water source area.(1)According to various national regulations and data,research and determine the types of risk sources of water sources.For the purpose of extracting various risk sources surrounding water source area from high-resolution remote sensing images,combining with the current situation of high-resolution remote sensing satellite data,we select the large scale land cover classification dataset as the foundational dataset so as to reduce the sensitivity of the water source risk source extraction model to time and location,and the dataset is modified to make the water source risk source extraction dataset.(2)Since Deep Labv3+ can extract multi-scale features of images,the paper uses it to extract the risk source surrounding water source areas from high-resolution remote sensing images.However,the extraction result will miss some objects with small areas and the object boundaries in the results is not consistent with real ground objects.In order to solve the problems,this article improves the Deep Labv3+ network from the two aspects of encoder and decoder.In the encoder,a feature extraction network with a channel attention mechanism is used to weight the channel information to improve the effectiveness of the network’s feature extraction;and then without changing the structure of the original network decoder,the skip connection method is used to combine other different semantic features with shallow detail features to improve the network’s ability to jointly utilize shallow and deep information.The improved network can better cope with the changes in the scale of water sources and risk sources in remote sensing images,maintain the edges of the features in the extraction results,and improve the leakage of small-scale features.(3)According to the documents related to the degree of pollution of risk sources,the weight indicators of various risk sources are determined.And we use the GF-2image as the main data source to quantitatively evaluate the non-point source of the water source with the non-point source risk index.Taking the Xiong’an New District of Hebei Province as the demonstration area,a study on the extraction of risk sources and the quantitative evaluation of non-point sources was carried out to verify the effectiveness of the method in this paper.