Study On The Oil Spill Detection In SAR Image Based On Deep Learning

In recent years,oil spill accidents have occurred frequently,which has brought great losses to economy and ecological environment.It has become a research hotspot to adopt effective methods to detect oil spill.Synthetic Aperture Radar(SAR)is characterized by all-day,all-weather,high resolution,and large range,which provides an effective means for detecting oil spill.Deep learning is a machine learning method that has developed rapidly in recent years.It can establish models for oil spill detection by learning the characteristics of samples.Convolutional Neural Network(CNN)is a widely used model in deep learning and has achieved some success in object detection.Based on deep learning,oil spill in SAR images is detected in this paper.The research contents are summarized as follows:1)An experimental study was carried out on the classic oil spill detection algorithms,oil spill detection algorithms based on Support Vector Machine(SVM)and principal component analysis(PCA)method were implemented.The basic principle of SVM is studied and analyzed,a suitable kernel function for detecting oil spill is selected according to the characteristics of the samples,and a model is constructed to detect oil spill,the accuracy of this algorithm is over 66%;The basic principle of PCA was studied and analyzed,the main components of oil spill and water were extracted and the main components of the samples were compared to detect oil spill,the accuracy of this algorithm is over 80%.2)Research to optimize back propagation algorithm of CNN model was implemented.The characteristics of the back propagation gradient descent method in the CNN model were studied and analyzed.In order to overcome the limitation of the local optimal,the first-order partial derivative in back error propagation part was modified to the second-order partial derivative to find the residual error,the method for calculating the partial derivative in the gradient calculation section was replaced by the secondary partial derivative in this paper.Based on the characteristics of cost function and excitation function in CNN model,the formula was derived and the efficiency evaluation experiment was carried out.The results of the experiment were only ideal before training 1000 times.3)The oil spill detection algorithm combined with gray level co-occurrence matrix and CNN were implemented.The principle,construction process,training method and experimental data demand of CNN model were studied and analyzed.For characteristics of oil spill SAR images under different polarization modes,the characteristic values of gray level co-occurrence matrix of different samples were calculated,and a CNN detection model for multi-polarization oil spill SAR images was constructed,whose parameters include number and size of convolution kernel,number of network layers,etc.The SAR images of HH,VV,HV polarization mode in L band were selected,and the oil spill detection experiments were carried out.The accuracy of this algorithm is over 85%.4)An oil spill detection algorithm combined with PCAand CNN was explored and studied.The oil spill detection algorithm combined with gray level co-occurrence matrix and CNN in this paper has high accuracy,but the operation time is long,therefore,an oil spill detection algorithm combined with PCA and CNN was designed using the dimensionality reduction characteristic of the PCA method.The principal components of gray level co-occurrence matrix eigenvalues are extracted based on PCA,which effectively reduced the dimension of the eigenvector input to the CNN model.The detection efficiency of the algorithm was improved,but the detection accuracy was lower than that before the improvement.