Hyperspectral Images Classification Based On Convolutional Neural Network

Hyperspectral remote sensing technology is developing rapidly,and the resolution of remote sensing data is getting higher.So it can provide a large amount of data information and valuable resources for subsequent research and applications.Therefore,the effective use of hyperspectral data is essential.Hyperspectral image classification technology is a branch of the research on hyperspectral data,it can provides feature information for subsequent research in various aspects such as agriculture,ecology,and the environment.But the process of hyperspectral data classification has the problem of high data dimensionality,high data volume,and small tagging samples to resolve.In order to improve classification accuracy,this paper uses convolutional neural networks to improve the classification accuracy of hyperspectral images.Specific studies include the following:First,a new three-dimensional convolutional neural network structure is proposed.By improving the original convolutional neural network structure,using three-dimensional convolutional layers replace the downsampling layer and the fully connected layer.Compared with most one-dimensional or two-dimensional convolution operations,three-dimensional convolution can better adapt to the three-dimensional characteristics of hyperspectral images than most operations using two-dimensional or one-dimensional convolution,avoiding the loss of information caused by the process of high dimensions reduce.This paper examines the feasibility of the convolutional neural network structure from the three aspects of training sample size,input data size,and network depth.The classification process is effectively completed without any preprocessing and post-processing.Experiments were performed on a variety of data sets,and the experimental results show that the structure can effectively improve the classification accuracy.Then,aim to the problem of data redundancy.Using some preprocessing operations can help extract effective information.However,many pre-processing operations are performed separately,which inevitably complicates the operation process.Therefore,end-to-end classification can be realized while reducing redundancy.This paper proposes to add a bottleneck attention mechanism based on the previous network,which can extract effective spectral and spatial information and reduce information.Then improve accuracy.Similarly,experiments have proved that adding this mechanism is effective.Finally,for the two methods proposed,a classification software is designed at the end of the article.Using Eric integrated development tools and Py Qt to create GUI applications,the software development of hyperspectral image classification method based on three-dimensional fully convolutional neural network and hyperspectral image classification method with attention mechanism is realized,also add some typical algorithm implementation.It combines multiple data sets and multiple parameter selections.The software can intuitively display the classification effect,which is very helpful for research and analysis of experimental effects.