Research On Target Detection Method Of Hyperspectral Image Of River Oil Spill Based On Deep Learning

Oil spill will cause serious pollution and ecological harm to the river environment.The detection of oil spill area is an important content of emergency treatment and protection of river environment.Hyperspectral image has the spatial and spectral characteristics of ground objects,and has been widely used in hydrology,agriculture,geological mapping and survey,environmental monitoring and other fields.However,in the detection of river oil spill targets based on hyperspectral image,it has become a new problem in its application because of its insufficient number of pixels,data redundancy,lack of sample number,affected by light and climate.The author introduces the depth learning theory into the river oil spill detection,and studies the target detection method of river oil spill hyperspectral image based on depth learning,in order to provide new ideas for solving the above problems.The specific work of this paper is as follows:1.Research on preprocessing and feature analysis of hyperspectral images of river oil spill.The simulation experiment of river oil spill is built,and the UAV with high spectrometer is used to collect the data of river oil spill.A method combining wavelet transform and principal component analysis is proposed to solve the influence of high-frequency noise such as environmental noise,equipment noise and solar reflection on river oil spill data;Solve the problem of oil spill hyperspectral image data redundancy.At the same time,the obtained river oil spill image is enhanced to provide more version changing images and realize the sample expansion of the original data.2.Based on spectral information,spatial information and spatial spectral information,CNN and DBN target detection models are built.According to the characteristics of river oil spill image,the primary spatial and spectral features are fused first,and then input into the established CNN and DBN target detection model to complete the image target classification.Through experimental verification,the two models successfully realize the regional target detection of crude oil,gasoline,engine oil and diesel in the hyperspectral image of river oil spill.The CNN model used in the detection results is better than the DBN model.3.An improved fast RCNN is constructed and applied to target detection in hyperspectral image of river oil spill.Although the CNN and DBN models have successfully completed the oil species identification in the oil spill area,the range and information of oil species in the oil spill area have not been accurately framed and detected,and the target framing and recognition can be achieved based on fast RCNN(fast region based revolutionary network),but it is rarely reported in the target detection of hyperspectral images.Therefore,combined with the characteristics of hyperspectral images of river oil spill,this paper improves the original network structure.Experiments show that the improved model has better classification effect than fast RCNN.4.A target detection model with double branch network structure is constructed.Aiming at the problem that 2D CNN model can not obtain spatial and spectral features at the same time,and 3D CNN model can not accurately capture spatial information,a double modal convolution neural network(dm-cnn)is proposed to obtain two-dimensional spatial features and three-dimensional features in the hyperspectral image of river oil spill at the same time.The experimental results show that the feature extraction efficiency of double branch convolution neural network in oil spill hyperspectral image is significantly improved.Compared with the traditional single branch network structure,the classification accuracy is the best.