Research On Remote Sensing Monitoring Of Polluted Water Bodies In Sea Estuaries And Upstream Rivers

In order to respond to the overall planning of land and sea proposed in the 14th Five-Year Plan,effectively guarantee the water quality safety of rivers entering the sea,protect and improve the marine ecological environment,it is urgent to carry out real-time dynamic monitoring of polluted water bodies in estuaries and upstream rivers.However,due to the complex water conditions in coastal waters,it is difficult to monitor the spatial and temporal variability of polluted water in large scale by traditional methods;Besides,the current mainstream identification models for black and odourous water bodies are sometimes difficult to extract them due to their strong regional characteristics,possessing poor universality for polluted water bodies and unstable identification accuracy.Therefore,in order to overcome these problems,this paper established a extraction model of sewage outfall water along the coast of Dongtai City,Jiangsu Province,based on Rayleigh-corrected reflectance(R_rc)data of Sentinel-2 satellite,which is expected to achieve automatic long-term real-time monitoring of non-nearshore urban polluted water bodies and near-shore sewage water bodies,in addition,established a recognition model of polluted water in Jiangsu Province based on machine learning algorithm,using R_rc data of GF-2 satellite.The main findings are listed as follows:(1)Pre-processing procedure of GF-2 satellite data including decompression,orthographical correction,radiometric calibration,image fusion and Raleigh correction,were automated and batched,which greatly avoids the loss of human and material resources caused by the heavy workload of data processing in the early stage,and improves the overall work efficiency.The processing results were similar to the results of mainstream products,which proved that they are credible.(2)Taking the estuary of some coastal shoal areas in Dongtai City,Jiangsu Province as the research object,the spectral difference between the sewage outfall water and other ground objects was analyzed by using Sentinel-2 R_rc data,then a decision tree identification model of the sewage outfall water bodies was established based on ratio index and chromaticity method.Satellite data from 2017 to 2020 were used to analyze the monthly distribution characteristics of polluted water bodies in the region,and it was found that the discharges were the least in January,August and September,while the higher emissions were in June,October and November.Most of the outfall water flows northward after entering the sea.In summer,it flows northward under the influence of southeast monsoon.While in spring,the northward flow trend is influenced by the northward tidal stress caused by the tidal wave extending from the Changjiang Estuary.(3)Based on GF-2 data,rich and representative images of urban water bodies were selected,and the morphological features of water bodies were highlighted by image enhancement and then assembled into a training sample library for training and validating U-net models,which is a deep learning algorithm.After iterative training,the model was highly accurate and stable.After that,images were used to verify the effect of the model and compare U-net with the traditional index method,and it was found that the traditional method has a poor extraction effect,while the U-net model has a high recognition accuracy except for a small part of building shadows.Next,the application and analysis of the model were carried out,and it showed that even if the background features were complex and there were many interference items,the U-net model could extract the water area accurately.(4)The spectral characteristics of GF-2R_rc data were used to analyze the differences between normal water and urban polluted water,and a rich and reliable sample spectral library was collected.Neural network,random forest and support vector machine,which are commonly used in machine learning algorithms were used to train and verify the identification model of polluted water bodies respectively.The results showed that the performance of the random forest model was more stable,not only the extraction results of polluted water were more consistent with the actual situation,but also there were no more misjudgment of normal water bodies such as ponds and irrigated fields.