Research On Deep Neural Network Based Classification Using Msi/Hsi And Elevation Data

Multispectral and hyperspectral data which contain abundant sepctral information have obtained great attention in land-cover classification.However,multispectral and hyperspectral images(MSI/HSI)do not perform well in distinguishing classes with similar properties in spectral bands.Digital models of a terrain's surface like digital surface models(DSMs)contain accurate elevation information,while they can't provide sufficient information for fine-grained classification.Spectral and elevation data have complementarities and their joint usage may open a new window for further improving classification accurarcies.However,differences between the two kinds of data make their effective joint classification a challenge task.This paper focus on feature level joint usage and classification of spectral and elevation data.A joint classification model for MSI/HSI and DSM data based on deep neural networks(DNNs)is designed.The constructed DNN-based models have powerful abilities in learning representative features and can further improve classicifation accurarcies for the better usage of heterogeneous information.Research work in this paper mainly includes the following aspects:Firstly,a feature extraction method based on convolutional neural networks(CNNs)is explored to extract invariant features from MSI/HSI and DSM data.Considering the importance of spatial information for land-cover classification,we choose CNNs,which are powerful in exploring spatially local correlation of data,to extract spatial-spectral information in MSI/HSI and spatial-elevation information in DSM data.Experimental results show that the deep feature extraction model outperform several widely-used feature extraction methods in remote sensing for classification.Secondly,a fully connected(FC)network is applied for the joint classification of spectral and elevation features.We stack the two features extracted from MSI/HSI and DSM data together to form a new feature vector which contains spatial,spectral and elevation information.A multi-layer FC network is then designed to perform nonlinear tranformation of the vector and accomplish the classification.Through experiments it can be seen that multi-layer networks have better performances than shallow models like support vector machines in classifying joint features.Finally,joint classification using DNNs and sparse representation is studied.Feature extraction is conducted on outputs of FC networks using dictionary learningand sparse coding to get sparse coding features which are less redundant and more discriminative.The extracted sparse coding features are then used as inputs of SVMs for the final classification and more accurarte results can be achieved.