Oil Spill Hyperspectral Remote Sensing Detection Based on DCNN with Multi-Scale Features

In this paper, a deep convolutional neural network (DCNN) model is developed for sea surface oil spill accurate detection using multi-scale features with AISA+ airborne hyperspectral remote sensing image. Based on multi-scale features after wavelet transform (WT), a deep convolution neural network classification algorithm with seven-layer network structure is proposed to detect oil spill of the Penglai 19-3C platform in 2011, and the accuracy evaluation is conducted on the overall situation. The detection results of proposed method are compared with those of the classical SVM, RF and DBN method. The results show that the accuracies of DCNN for oil spill detection based on different-scale features are all more than 85 %, which are much better than those of SVM, RF and DBN method, and the detection results can maintain the continuity of oil film at sea. Among them, the detection result of DCNN model based on spectral feature information combined with low-frequency component of 1-level wavelet transform has the best effect and highest detection accuracy, reaching 87.51 %.