Hyperspectral Image Classification Based On Lightweight Network And Incremental Learning

Hyperspectral image Classification has attracted significant attention in the field of earth observation.In the era of big data,the amount of data obtained by advanced remote sensing sensors are ever-increasing.With the expansion of application scenarios and the increasing application requirements,new data categories and fine categories appear from time to time.Faced with such an open and dynamic environment,class incremental learning,a fundamental solution,has received huge attention.Recurrent neural networks,convolutional neural networks,and other networks for hyperspectral classification usually perform well in this field,but they require too much computing resources and time.Therefore,the primary purpose of this thesis is to study fast and lightweight deep learning algorithms that can efficiently implement the incremental classification of hyperspectral images.The main work is as follows:1)This paper proposes a linear programming-based incremental learning(LPILC)method that can learn a new category from a limited number of samples through a few iterative steps.This method uses a linear programming model to modify the classifier weights so that the incremental part can adapt to the new category data while benefiting from the pre-trained model.This method can largely reduce the computational cost and time cost while adapting to new HSI data.This article applies the proposed LPILC to two scenarios of hyperspectral image new class and refined classification and considers four data situations to illustrate the superiority and robustness of the proposed LPILC in the case of insufficient data.A large number of experiments have shown that LPILC can achieve the highest hyperspectral classification accuracy in the shortest time under the same dataset and computing resources.2)In this paper,considering the fast response performance of the echo state network,a multiscale learning hyperspectral classification method of the adaptive leakage rate echo state multilayer perceptron is designed.This method solves the problem that the traditional echo state network cannot adaptively modify the leakage rate,resulting in poor classification near the target edge.Based on that,an echo state network combined with multilayer perceptrons and multi-scale learning strategies is proposed.It greatly reduces the training complexity and training time while ensuring the overall classification accuracy.At the same time,this chapter explores the effect of different adaptive leak rates on data sets with small homogeneous regions.Experiments show that the adaptive leak rate function proposed in this paper can retain the edge information well and lead to higher accuracy.3)In the traditional convolutional neural network,the feature extractor will change the feature space when performing incremental classification,leading to the change of the features of the original category.Based on this,this paper proposes a feature extractor based on echo state that can achieve incremental learning while maintaining the original category features.The feature extractor based on the echo state proposed in this paper can extract various samples of hyperspectral images into a feature space with good feature dispersion.Through the visualization results and classification accuracy of the echo state matrix and trajectory,it can be seen that the features extracted by the feature extractor proposed in this chapter have large differences between classes,the classification effect is good,and it has continuous incremental classification capabilities.