Deep Learning Method And Software Implementation For Fusion Classification Of Hyperspectral Image And LiDAR Data

The hyperspectral image(HSI)is composed of hundreds of continuous spectral channels and contains rich spectral and spatial information.The LiDAR data covers high-precision three-dimensional spatial information such as the altitude,shape and scale of the surveyed area,which can be used as an additional supplement to HSI,effectively supplementing its feature diversity in the classification process.In recent years,the fusion classification of HSI and LiDAR data has gradually developed into a key technology for remote sensing data analysis.This paper focuses on deep learning-based HSI classification and fusion classification of HSI and LiDAR data.The main contributions are as follows:(1)For the problem of HSI supervised classification,a deep learning classification algorithm for HSI driven by partial differential equations is proposed.The PeronaMalik(PM)diffusion equation is introduced into the hidden layer of convolutional neural network(CNN),which not only retains the feature extraction ability of CNN in deep network,but also introduces the edge preservation and noise reduction of PM diffusion.Then 1×1 Convolution and this novel convolution are combined to form a TPM-block,and an end-to-end deep learning network is constructed by cascading multiple TPM-blocks.The experimental results show that the algorithm not only enhances the deep feature learning ability of CNN,but also improves the ability of preserving the boundary of ground objects in the process of supervised classification.(2)Aiming at the fusion classification problem of HSI and LiDAR data,a deep learning algorithm driven by diffusion equation is proposed.The algorithm fully learns the deep features of the two types of remote sensing data by cascading multiple convolutional layers and trainable PM diffusion units;at the same time,we also proposes a novel probability reconstruction fusion algorithm,which first combines the heterogeneous features and the fusion features obtained by feature summation are input into softmax classifier to obtain their respective probability matrix.And then the probability matrices are re-integrated through weighted summation to generate the final classification probability prediction.Experiments show that the algorithm has excellent classification performance.(3)We propose a deep learning algorithm with a dual-branch multi-level fusion structure to solve the problem of imbalance between heterogeneous features during the fusion of HSI and LiDAR data.The algorithm uses a dual-branch structure to learn the deep feature information of HSI and LiDAR data respectively.At the same time,in order to fully integrate the heterogeneous features of these two kinds of remote sensing data,feature-level fusion and decision-level fusion are established.Among them,in order to solve the problems of a large amount of parameters and over-fitting in the process of fusion,the widely used fully connected fusion strategy is replaced by the feature summation.For the decision-level fusion,a weighted summation strategy is adopted,where the weights are calculated from the classification result of each feature.Experiments show that the proposed algorithm has excellent classification performance.(4)Design an integrated system of classification algorithms.The system covers the two classification algorithms proposed in this paper and the other six comparison algorithms,which integrates visualization,classification and classification performance evaluation.It achieves fast,convenient and accurate capture of feature information from input data and obtains classification results.