Research On Hyperspectral Image Open-set Classification Based On Convolutional Neural Network

Owing to the rapid development of hyperspectral remote sensing technology,pixel-level classification of hyperspectral remote sensing images has become a research hotspot in the field of remote sensing.In recent years,classification algorithms based on convolutional neural networks have greatly promoted the development of classification techniques for hyperspectral images(HSI).However,the trend is to improve the classification accuracy by continuously deepening the network layers or increasing the complexity of the model structure,which brings about high time consumption and hinders the application of classification algorithms.Furthermore,the existing classification models assume the environment of HSI classification as closeset scenario,and classify all pixels into known categories that appeared during training.They ignore unknown samples that are ubiquitous in the real situation,which decreases actual classification accuracy in practical applications.This paper aims at the shortcomings of the existing close-set classification algorithm for HIS.Combining with open-set classification methods in computer vision and natural language processing,the classification algorithm for HSI is improved.We use standard datasets such as Indian Pines,Pavia University and Kennedy Space Center for experiment.The details are as follows:1.The development status of HSI classification and open-set classification is analyzed.Moreover,combining the data characteristics of HSI data and the definition of open-set classification,the basic concepts and evaluation criteria of open-set classification for HSI are introduced.2.Aiming at the high time consumption of high-precision close-set classification algorithm,a fast convolutional neural network for continuous spatial-spectral feature extraction(FSS-CNN)is proposed as the basic classification framework of open-set classification algorithm.To achieve the purpose of reducing network parameters,we adjust the spatial-spectral feature extraction method by decomposing the threedimensional convolution kernel,the network is divided into the 1×1 convolution stage to reduce spectral dimension and the ordinary two-dimensional convolution stage to extract spatial feature.The experimental results show that the method improves the calculation speed while maintaining high accuracy.3.Aiming at the problem that unknown samples in the close-set classification algorithm are misclassified into known classes,an open-set classification method based on convolutional neural networks and probability thresholds(OS-CNPT)is proposed,and the construction of open-set dataset for HSI is given.To achieve the purpose of identifying unknown samples,we set threshold for the classification probability output by the Soft Max layer in FSS-CNN.In the threshold optimization stage,we use the proposed index which balances the overall accuracy of the known classes and the unknown class recall as the optimization index.The experimental results prove that the algorithm has obvious advantages in open space risk,visualization results and various open-set classification evaluation indices,which verifies the rationality of the balance score.