Hyperspectral Data Haze Monitoring Based On Deep Learning

Automatic haze monitoring is one of the key issues of environmental governance,because the accuracy of the remote sensing haze monitoring is poor and the cost of the ground haze monitoring is very high.To solve those issues,two novel haze monitoring methods are proposed,which combine deep learning network with hyperspectral data.The main works of this thesis are as follows.1)Physical properties and spectral characteristics of haze are analyzed.The physical characteristics such as composition,humidity and particle size of haze are studied.According to the research of Hyperion data,we make the conclusion that the radiation intensity received by the satellite sensor of haze day is large than non-haze day.But the actual spectral curve is more complex,we need to solve this problem by deep learning.2)A novel haze monitoring method based on deep blief network(DBN)is proposed,which can make the whole network easier to train.The addition of dropout improves robustness of the model.The experimental results on Hyperion hyperspectral data in the Suzhou region show that the proposed DBN method has higher haze monitoring accuracy in comparison with SVM and traditional neural network algorithms.3)Based on deep residual network(DRN),a novel haze monitoring method is proposed,which combines deep learning network with hyperspectral data.First the features of haze hyperspectral curves are obtained by deep network,then the network is optimized using residual leaning,eventually haze monitoring model is achieved.The experimental results on Hyperion hyperspectral data in the Suzhou region show that the proposed DRN method has higher haze monitoring accuracy in comparison with state-of-the-art methods.4)Built a haze monitoring system with GUI for PC.The system includes modules of data display,data processing,evaluation and output.