Identification Of Hyperspectral Mineral Type Using Deep Learning Method Supported By Ground Object Spectral Library

Hyperspectral remote sensing technology plays an important role in mineral mapping.The identification of hyperspectral mineral types is a key step to achieve mineral map.The complexity of mineral types and the environment makes the traditional method of mineral recognition not accurate enough to meet the application requirements of mineral mapping.Deep learning technology developing in recent years has a strong ability to learn the essential characteristics of the sample data set.The technology is increasingly used in the analysis of hyperspectral data.While using neural network of deep learning technology to identify mineral types of hyperspectral data,the focus is to extract spectral characteristics of minerals from the sample data.Making a label sample data set by collecting sample points from hyperspectral remote sensing data is time-consuming,inefficient,and difficult to provide a sufficient number of effective samplcs.On the other hand,the quality of network sample data is affected by factors such as the quality of remote sensing data and the accuracy of preprocessing.Therefore,the insufficient quantity and low quality of the sample data make the application of deep learning restricted in the identification of hyperspectral mineral types.Aiming at the above problems,this paper takes the existing data of the ground object spectral library as the effective input sample data of the deep learning network.And a hyperspectral mineral types identification method using deep learning based on the ground object spectral library is proposed.The main contents of this article include:(1)Production for deep learning mineral type identification samples based on USGS spectrum library.The training samples in the deep learning algorithm determine the prediction accuracy to a certain extent.The existing ground feature spectral library is selected as the sample data set,the processing of sample data is omitted,and the mineral recognition efficiency of the algorithm is improved.There are many mineral types and amounts of spectral data in the USGS spectral library to meet the requirements of deep learning samples.At the same time,the data information of the spectral library has high accuracy,which can more realistically reflect the spectral characteristics of minerals,and meet the quality requirements of deep learning samples.A large number of label samples for deep learning mineral type recognition are produced based on simulation methods by using the USGS spectrum library and provide an effective spectral data set for the input layer of the network.(2)Deep learning network structure design and hyperspectral mineral identification method.Aiming at the characteristics that the spectral difference of mineral types is relatively small and the interference information is significant,this paper selects the residual network model to realize the identification of mineral types.As a typical convolutional neural network,the residual network has a simple network structure,few parameters,and strong generalization.When processing spectral information with large interference and correlation in hyperspectral data,a stronger recognition effect can be achieved.Based on the Keras deep learning framework platform,this paper constructs a residual network for mineral types identification of hyperspectral data,and input the spectral sample data supported by the USGS spectral database into the network.The residual network extracts the spectral characteristics of the data by the deep network structure,and learns the rules for identifying hyperspectral minerals based on those features.Using the AVIRIS data of Cuprite mining area as the data to be identified,the mapping experiment based on the mineral identification rules obtained by network training was carried out to obtain the mineral mapping results of the research area.(3)Authenticity test and uncertainty analysis of mineral type identification results.Based on the mineral mapping results combined with surface survey and remote sensing images,the mineral type recognition algorithm proposed in this paper is verified by comparative analysis with mineral mapping results of deep learning.The overall accuracy of the mineral identification algorithm reached 76.41%,and achieved a good mineral identification effect.This proves the feasibility and effectiveness of the deep learning method using the ground object spectral library as the sample data for the classification of hyperspectral minerals.(4)The adaptability analysis of mineral type identification in the vegetation area.In the mixed spectrum of remote sensing data in vegetation coverage area,the spectral information of vegetation affects the accuracy of the identification of hyperspectral mineral types.In order to verify the applicability of the mineral type identification method in the vegetation area,the mineral mapping experiment was carried out on the hyperspectral data with vegetation.The linear spectral mixing model is used to simulate the mixing spectral of vegetation and mineral to obtain simulated hyperspectral data of vegetation area.The data is used for mineral type identification experiments and analysis of the identification accuracy of the results.It proves that the deep learning method of ground feature spectrum database as sample data has high applicability and stability.