SAR Oil Spill Images Classification Based On CF-DBN In Small Sample Sets

Lookalikes(such as atmospheric gravity wave,ship trace,low wind speed area,biological oil film,leeward headland)and oil spills all present dark spot characteristics on SAR image,which is easy to be confused during identification,thus improving the false alarm rate of oil spill detection on the sea surface.In recent years,deep learning has made great progress in image classification.Successful deep learning applications often rely on sufficient sample sets,but the frequency of radar satellite transit is generally low,oil spills on the sea surface are highly polluting and need to be cleaned up in time.Therefore,the number of SAR oil spill image samples is often insufficient.Based on this,this paper focuses on the classification of SAR oil spill images based on small sample sets,and proposes an improved Deep Belief Network(DBN)——Cascaded Feed-forward Deep Belief Network(CF-DBN),and preliminarily verifies the effectiveness of this model through experiments.The specific work of this paper is as follows:(1)The five SAR images under the HH channel are pre-processed,including denoising,geometric correction,ROI extraction.Then the image is segmented by slider and cut into appropriate size to facilitate subsequent feature extraction.A small data set of 1500 samples(500 oil spill samples,500 lookalikes samples,and 500 seawater samples)was finally obtained.(2)Feature extraction of the data set,a total of 33 texture features of Tamura,gray level co-occurrence matrix(GLCM),gray-gradient co-occurrence matrix(GLGCM)and gray difference statistics(GLDS)were extracted.And the 10 features with better discrimination are selected to form the feature vector as the input of the model.(3)An improved DBN——CF-DBN is proposed.The improvement point is to replace the BP Neural Network(BPNN)in the traditional DBN model with a Cascaded Feed-forward Neural Network(CFNN).The CFNN has fast learning speed and better nonlinear fitting characteristics,which can effectively reduce the model training time and improve classification accuracy.CF-DBN is used to classify SAR oil spill images.At the same time,the same data set was applied to DBN,CNN and three other traditional machine learning methods(BPNN,SVM and CART)for comparison according to classification results.(4)According to the evaluation criteria such as accuracy rate,running time,Kappa coefficient and ROC curve,it is proved that CF-DBN is effective in SAR oil spill images classification with small sample sets.Although it takes more time than traditional machine learning algorithms such as BPNN,SVM and CART,the training speed is increased by about 10 times compared with deep learning algorithms such as DBN and CNN.CF-DBN can distinguish oil spills and lookalikes region on SAR image.