Intelligent Detection Of Target In Complex Nearshore Scenes Based On SAR Image

With the continued rapid development of the global economy and frequent human activities in coastal areas,more efforts are necessary to be devoted to regulating marine exploitation,fishing and other agricultural activities,for ocean ecological environment protection.To this end,the State Oceanic Administration has formulated the National Marine Functional Zoning Plan(2011～2020)to improve the comprehensive management of different ocean areas.However,it is very common to observe marine pollution,for instance,the operational and accidental discharges(oil spills,bulk toxic liquid substances,domestic sewage,ship garbage,harmful exhaust gas,etc.)from various ships during their voyages.In addition,the self-purification capacity of seawater is reduced,and the marine ecological balance is threatened due to extensive aquaculture and fishing activities.Therefore,accurate and rapid monitoring of coastal ships and aquaculture is necessary to overcome the above problems and optimize the navigation of ships as well as the layout of aquaculture space.Synthetic Aperture Radar(SAR)is of great help to provide continuous observation data information for accurate extraction and monitoring of ship activities and mariculture near-shore,benefiting from high resolution,high penetrating power,and weather-independent.Up to now,thanks to the rapid development of deep learning technology,it attracts much attention to applying the deep learning methodology to acquired SAR images.Some studies have shown that it has good performance in classifying the various objects in SAR images.Therefore,in this thesis,based on the present deep learning method,we exploit the potential to detect ships and floating raft aquaculture on satellite data(Sentinel-1 and the Gaofen-3(GF-3)).Thereinto,the scattering information from the rich feature,and statistical distribution characteristics of SAR images and SAR target imaging mechanism are considered in developing the deep learning method.The three parts of this thesis are:(1)A two-stage ship detection method based on a lightweight network is proposed in order to reduce high false alarms,high complexity,and interpretability of the current deep learning.The method first generates candidate targets from the preliminary screening of ships and false alarm targets according to the nearshore background suppression filter and the statistical distribution characteristics of radar data after morphological processing.Then the model network is simplified into two layers,in particular,the shallow layer to capture and learn different scale ship features.It effectively reduces the false alarm rate by distinguishing ship targets from complex At the same time,the Grad-CAM theory is used to strengthen the interpretability of the model for detecting near-shore ships with dual-polarizations(VH/VV)of Sentinel-1images.The results show that the data in VH polarization can suppress the azimuth ambiguity and the background objects(e.g.,islands)better than the VV polarization.Overall,this method effectively reduces the false alarm in the detection of ships on different scales and improves the detection accuracy.(2)In order to solve the problem of deep learning ship detection for complex nearshore scenes overly relying on a large number of reliable labeled samples,inefficient training,redundant information,and difficulty in using the variability between samples for feature learning from unlabeled data species,a two-stage near-shore ship detection method based on small-sample active learning is proposed.By introducing an active learning mechanism,optimizing the structure of the deep learning model,randomly selecting a small number of labeled samples,training the model to learn and express features from unlabeled samples,and using the uncertainty principle to automatically label unlabeled samples of reliable quality for iterative training.The experiments show that the required labor cost and time can be significantly reduced while maintaining high near-shore ship detection accuracy,enabling near-shore ship detection under small sample conditions.The experimental results show that in complex nearshore areas,the method in this paper can effectively reduce the number of labeled samples,and the detection accuracy of VH polarized nearshore ships with only 50 samples randomly labeled has an F1 coefficient of 0.96,and it is suitable for detecting different nearshore ships.(3)In order to solve the deep learning methods that ignore the guiding role of the target imaging mechanism in nearshore target detection,the prediction results are highly dependent on training data,a lightweight network extraction method for nearshore floating raft aquaculture that take into account polarization characteristics is proposed.The transparency of the deep learning algorithm is enhanced by adding interpretable target imaging mechanisms to the data-driven deep learning method,and then the necessary features are analyzed and determined from polarization and scattering features to extract floating raft aquaculture to reduce the transition dependence of the model on training samples.Finally,the model is optimized by reducing the model redundancy parameters through depthwise separable convolution and residual structure to improve the training efficiency and further validate the effectiveness of the necessary features guided by the imaging mechanism for extracting intertidal floating raft aquaculture in different backgrounds.The experimental results show that the optimal HV polarization feature obtained from the SAR imaging mechanism of the nearshore floating raft aquaculture contains the most abundant information.The accuracy in the optimized LU-Net model(HV: IoU=0.83)is better than the extraction accuracy of HH and VV polarization,and almost equal to the multifeature(HH+HV+VV: IoU=0.84),and the model training is easier to converge.