Remote Sensing Extraction Of Black-odorous Water Bodies Based On Deep Learning

Black-odorous water bodies are scattered and widely distributed in China,which has a serious impact on people’s life and health,social economy,urban beauty and so on.It is a common environmental pollution problem in many large and medium-sized cities in China.At present,the research on remote sensing identification methods of black and odorous water body at home and abroad mainly focuses on traditional methods,such as artificial identification method,chromaticity method,index method and so on.However,due to the complexity of remote sensing image background,terrain diversity,black-odorous water bodies and non-black-odorous water bodies on spectral information has certain similarity,and the black-odorous water bodies are mostly smallshaped canals,It is difficult to distinguish between black-odorous water bodies and nonblack-odorous water bodies and difficult to accurately identify the shortcomings and so on small target with traditional remote sensing method.(1)In order to improve the traditional remote sensing method to extract the disadvantages of black-odorous water bodies remote sensing information,using deep learning technology,build the PSPNet(Pyramid Scene Parsing Network),Deep Labv3+ and U-Net Network models of black-odorous water bodies remote sensing information extraction of contrast experiments,prove that deep learning technology can be well applied to black-odorous water bodies in the remote sensing information extraction;(2)In order to distinguish black-odorous water bodies and non-black-odorous water bodies,enrich the ground feature band information of the image and improve the extraction accuracy,the image band calculation is carried out to generate the derived band: NDVI(Normalized Difference Vegetation Index)and NDBWI(Normalized Difference Black-Odorous Water Index)fused the derived bands with the original image,In three kinds of deep learning network in the model input RGB three-channel,RGB + NIR four-channel image,RGB + NIR + NDVI five-channel image,RGB + NIR+ NDBWI five-channel image,RGB + NIR + NDVI + NDBWI six-channel image,a total of five kinds of combination image comparison,the study found that NIR+RGB+NDVI+NDBWI six-channel image with the most abundant characteristic information and less data redundancy,in every model can better distinguish blackodorous water bodies and non-black-odorous water bodies,improve the extraction accuracy,However,due to the complexity of the background features of remote sensing images and the small shape of black-odorous water bodies,there are still problems of missed detection and wrong detection;(3)In view of the small shape black-odorous water bodies extraction of residual and the fault detection problem,adding the attention mechanism module into the model,the channel attention and spatial attention two aspects to optimize the model,can be a good help model learning and black-odorous water bodies related pixels,inhibition models of learning has nothing to do with the black-odorous water bodies pixels,shape of small black-odorous water bodies extraction difficult problem was improved obviously,and the introduction of attention mechanism module model extraction effect than the same input channel combination of the original model better extraction effect,among them,Input RGB + NIR + NDVI + NDBWI six-channel remote sensing image and introduce attention mechanism of U-Net network model best,the result of the black-odorous water bodies extract can accurately distinguish between black-odorous water bodies and non-black-odorous water bodies,and to the shape of tiny black-odorous water bodies extract,improve the residual black-odorous water bodies remote sensing information extraction and fault detection problems,greatly improved the extraction accuracy,the accuracy of the evaluation index of F1-srore,MIo U,Recall reached0.8645,0.8681,0.8359.