Multi-channel SAR Image Classification Based On Deep Learning And Its Application In The Environmental Monitoring Of The Intertidal Zone Of Shanghai Nanhui Dongtan

With the development of satellite remote sensing technology,the interpretation of remote sensing images has shifted from target discrimination to pixel-level classification.Synthetic Aperture Radar(SAR)is one of the important remote sensing data sources for image interpretation.Its ability to work continuously under all-weather and all-weather conditions has been widely used in military and civilian fields.Application,classification technology requirements for SAR images are getting higher and higher.With the vigorous development of computer technology,Deep Convolutional Neural Network(DCNN)methods represented by deep learning have emerged one after another.Compared with traditional image classification methods,it has achieved good performance in optical image classification tasks due to its features such as advanced context extraction,end-to-end learning methods,and high classification accuracy.However,due to the difference in imaging mechanism between SAR image and optical image,DCNN is difficult to interpret SAR image.This study proposes a new DCNN model for multi-channel SAR images to improve deep learning models'performance in multi-source remote sensing information fusion.Besides,experiments have verified the effectiveness and advancement of the model.The main research contents of this paper are as follows:(1)A special dataset for environmental monitoring of Shanghai Nanhui Dongtan's intertidal zone was constructed.This dataset uses SAR dual-polarization channel and physical environment information in Shanghai Nanhui Dongtan's intertidal zone from2016 to 2020 as the data source.Combining prior knowledge of experts,multi-source remote sensing data to assist visual interpretation and other means,Divide data into five categories and label them as label images.Provide research data support for this article and subsequent research work.(2)For multi-channel SAR image classification,explore the classification effect of dual-polarization SAR image fusion of multi-source information.Based on pre-processed VV and VH single-polarization datasets,we successively constructed VV+VH dual-polarization dataset and VV+VH+W+T dataset integrating tide level and wind speed information.We Used the U~2-Net model to conduct a combined comparison experiment to verify multi-source remote sensing information fusion effectiveness.(3)A multi-branch DCNN model(MB-U~2-ACNet)with asymmetric convolution for U~2-Net framework is proposed.This model has a dual-layer nested U-shaped encoder/decoder structure.We extract and fuse characteristics of multi-source remote sensing information by establishing a branch structure.Simultaneously,we also designed a new Asymmetric Convolution Re Sidual U-block(ACRSU)for each level of encoder/decoder to enhance the network's ability to resolve image flipping and rotation.Besides,by introducing an attention mechanism to distinguish the importance of features from the perspective of the channel and spatial dimensions,the model's performance can be improved.(4)Based on the dataset constructed in this study,an image sequence with a time interval of less than one month between adjacent images is established through manual screening.In two dimensions of time and space,the study area's typical vegetation features from 2016 to 2020 are analyzed from overall,seasonal,and inter-year forces,and driving forces that affect vegetation coverage change results are analyzed.This study verifies the effectiveness of multi-source remote sensing information fusion on the constructed Shanghai Nanhui Dongtan intertidal zone monitoring dataset.Moreover,through multiple sets of experiments,it is proved that MB-U~2-ACNet proposed in this study has better performance than existing representative DCNN and traditional methods.Simultaneously,temporal and spatial characteristics of typical vegetation features in the study area are analyzed.The research results obtained in the thesis can be used for long-term environmental monitoring of Shanghai Nanhui Dongtan's intertidal zone,providing theoretical and data support for the construction of the intertidal zone monitoring system,resource,and environmental assessment and policy formulation.