Hyperspectral Target Detection Based on Sample Enhancement and Automatic Parameter Optimization
Hyperspectral target detection based on deep learning faces challenges such as insufficient quality of samples,intri-cate network structures,and laborious parameter adjustment.In this paper,we propose a deep learning method with data augmenta-tion and automatic hyperparameter optimization.To tackle the issue of insufficient quality of samples,we introduce a sample aug-mentation strategy.The strategy utilizes endmember extraction and clustering techniques to directly acquire a large number of back-ground pixels from hyperspectral images.By pairing these with a small number of known target pixels using a phase-reducing pixel pairing approach,we obtain a large number of labeled pure sample pairs,thereby accomplishing data augmentation.In addition,dis-tinct from most complex deep networks,we designed a lightweight Convolutional Neural Network(CNN)comprised of 12 convolu-tional layers.This network is specifically engineered to efficiently and rapidly learn the mapping between input sample pairs and their corresponding labels.By incorporating the particle swarm optimization algorithm,this network possesses the capability to auto-matically optimize hyperparameters,overcoming the shortcomings of laborious parameter adjustment.This enables the network to automatically adjust hyperparameters based on samples from different hyperspectral images,thereby generating optimal results.For a test pixel,the input to the trained network is the spectral difference between the central pixel and its adjacent pixels.When a test pixel belongs to the target,the output score is closely align with the target label.Experimental results on five hyperspectral datasets demonstrate that our method significantly outperforms existing techniques.
NETL
NSTL
万方数据
