Research On Water Body Recognition And Water Quality Inversion Of Remote Sensing Images Based On Deep Learning

Due to human activities and climate change,surface water resources are facing varying degrees of threats,and issues related to water coverage area and water quality have become increasingly prominent.Timely and effective monitoring of water coverage area and water quality parameters is the foundation for the rational development and utilization of water resources,as well as for the harmonious development of society and ecology.In recent years,remote sensing technology has been widely applied in the extraction of water surface and water quality monitoring.Remote sensing technology has the advantages of rich spectral bands and wide coverage range,which can quickly,repeatedly,and accurately obtain water information and monitor surface water resources.In this paper,we focus on surface water environmental issues,and use GF-2 and Landsat-8 to respectively establish water body segmentation and Poyang Lake chlorophyll-a inversion data sets,and explore more accurate deep learning-based water body segmentation and chlorophyll-a inversion algorithms.The specific research contents and results are as follows:(1)For deep learning-based remote sensing image water body segmentation and chlorophyll-a inversion tasks,we have created GF-2 remote sensing image water body segmentation data set and Landsat-8 chlorophyll-a concentration inversion data set.(2)In order to explore higher accuracy remote sensing image water body segmentation methods,several classical backbone feature extraction networks and segmentation models were compared for their accuracy on the GF-2 water body segmentation dataset.Among them,the U-Net model with VGG16 as the feature extraction network achieved the highest MIoU,MPA,and Acc,reaching 95.24%,97.27%,and 99.12%,respectively.Secondly,to avoid the loss of local information caused by global pooling in attention mechanisms,a multi-scale fusion attention module(MSFAM)with depth-wise separable convolution was designed and used to enhance effective feature information after U-Net skip connections.Finally,dilated convolution was used to increase the receptive field in the VGG16 feature extraction network.After these improvements,the water body segmentation accuracy was further improved,with MIoU increasing by 1.04%compared to U-Net,and MPA and Acc reaching 97.98%and 99.26%,respectively.(3)A one-dimensional convolution-based model for chlorophyll-a concentration inversion in Landsat-8 was constructed.In order to achieve higher inversion accuracy,comparative experiments were conducted with different convolution kernels,fully connected layers,and optimizers.The experimental results showed that the one-dimensional convolution model with dilated convolution and Adam optimizer achieved the highest determination coefficient of 0.92 and the lowest RMSE of 0.44.Subsequently,a heat map was created and eutrophication evaluation was performed for chlorophyll-a concentration in Poyang Lake,further verifying the superiority and practicality of the one-dimensional convolution model.