Research on SAR oil spill image classification based on DBN in small sample space

SAR has become one of the important means of oil spill monitoring. However, oil spills and lookalikes are characterized by dark spots on SAR images. They have similar or identical backscattering coefficients and gray values, which are easy to produce confusion. Aiming at this problem, this paper proposes a deep learning model-Deep Belief Network (DBN), which uses DBN to distinguish oil spills, lookalikes and water. In the experiment, 900 images were collected from the three SAR oil spill images to form a small sample space dataset. The two kinds of texture features such as Tamura and Gray Level-Gradient Co-occurrence Matrix are extracted, and the feature vector with good distinguishing features is selected as the input data of the model. Finally, the classification results are compared with the traditional machine learning method (BP, SVM). The experimental results shows that the DBN model proposed in this paper is superior to these classifiers in classification accuracy.