Research On Hyperspectral Image Classification Algorithm Based On 3D Convolutional Neural Network

Hyperspectral remote sensing technology is a significant promising technology for research.Hyperspectral images can contain a large amount of spatial-spectral information,reflecting to some extent the physical and chemical properties of the photographed object.Therefore,it has important research value in the fields of environmental testing,geological analysis,agricultural monitoring,atmospheric monitoring and so on.Hyperspectral image classification,as a branch within the field of hyperspectral image research,plays an important role for research in various fields.The use of computational devices to classify the pixel of hyperspectral images can save a lot of labor,material,and time costs.When facing the high-dimensional characteristics of hyperspectral images,the general approach is to first reduce the dimensionality of hyperspectral images before classifying them using classification models.However,there is no guarantee that the reduced-dimensional image will necessarily filter out all the redundant spectra,leading to the limited effectiveness of all the general classification models.To better improve the effectiveness of the classification model and to provide an effective idea to face such problems in the future.This paper focuses on the classification method,combined with 3D convolutional neural network,to investigate the hyperspectral image classification problem.The main research of this paper is as follows.(1)First,this paper provides a relatively detailed introduction to the background characteristics of hyperspectral images and the current domestic and international research status.After that,for the common or prospective methods of hyperspectral image processing,this paper introduces one of the algorithms selected from different research fields in hyperspectral image classification: Principal component analysis,Deep reinforcement Learning and SE module.For classification models,support vector machines and convolutional neural networks are chosen to be introduced in this paper.Finally,a more detailed description of common evaluation metrics in hyperspectral image classification is presented.(2)Due to the numerous spectral dimensions of hyperspectral images,most classification models suffer from issues such as breaking spectral continuity and poor learning of spectral information.In this paper,we propose a new classification model called the enhanced spectral fusion network(ESFNet),which contains two parts: an optimized multi-scale fused spectral attention module(FsSE)and a 3D convolutional neural network based on the fusion of different spectral strides(SSFCNN).Specifically,after sampling the hyperspectral images,our model first implements the weighting of the spectral information through the Fs SE module to obtain spectral data with a higher degree of information richness.Then,the weighted spectral data are fed into the SSFCNN to realize the effective learning of spectral features.The new model can maximize the retention of spectral continuity and enhance the spectral information while being able to better utilize the enhanced information to improve the model’s ability to learn hyperspectral image features,thus improving the classification accuracy of the model.Through many experiments for analysis,the model in this paper obtained 90.1% and 96.1% accuracy on the Indian Pines dataset and the Pavia University dataset,respectively,all of which are the highest accuracy,and it has a good classification effect.(3)Aiming at the high-dimensional and low-resolution characteristics of hyperspectral images,in order that the model can learn the spectral features of hyperspectral images more effectively.This paper continues to design and propose a novel hyperspectral image classification model,the Spatial-Spectral Pyramid Network(SSPN),using 3D convolutional neural networks.The model can obtain many spectral features with different levels of abstraction by multi-scale convolutional extraction.The feature maps at different scales are then fused in a two-by-two fusion to further enrich the features contained in a single feature map.It allows the model to better learn the features between spectra at different scales of hyperspectral images and to complement the feature maps at different scales.Finally,by weighting and summing the losses calculated after the combination,the model can avoid a single combination that overly affects the update of the model parameters.It will make the design of the model in this paper more effective and reasonable.Through many experiments for analysis,the model in this paper achieved 99.0%,99.8%,100% and 98.7% accuracy on the Pavia University dataset,Botswana dataset,Chikusei dataset and Houston2013 dataset,respectively,all of which are the highest accuracy,and it has a good classification effect.