Research On Hyperspectral Imaging Technology Based On Compressed Sensing

As a remote sensing technology gathering spatial information and spectral information at the same time,hyperspectral imaging is one of the important detection techniques in the field of modern remote sensing.Traditional hyperspectral imaging technologies comprise the techniques such as point-FOV scan,line-FOV scan,and plane-FOV staring for acquisition.These techniques are based on the Nyquist sampling theory for data acquisition;to ensure the information is accurately retained,the sampling frequency must be greater than twice the maximum frequency contained in the detection signal;moreover,it's necessary to finally guess the outcome data after data compression and decompression.This information processing process will bring a huge waste of computing space and storage space for the following reasons: The system needs to store the acquired data containing a lot of redundant information;the data is compressed by complex compaction algorithms to attain very little available data;undoubtedly,this posed a huge challenge for some systems with limited computing and storage capabilities such as the on-satellite system.Compressive sensing is an information processing theory that achieves compression and sampling at the same time,and can thus overcome the conflicts caused by the above-noted process;the compressive sensing-based spectral imaging system has also become one of the current focuses of research.This paper includes the research contents with the following innovations based on the study on the compressive sensing-based imaging mechanism:1.The slit group coded scanning hyperspectral imager(SGCSHI)technique was proposed.Slit coded based on the compressive sensing theory,SGCSHI is designed to get all information in the event of under-sampling;experiments have shown that the quality of spectral image achieved with SGCSHI is equivalent to that achieved with traditional single-slit sweep,but the detection efficiency(frame rate)is increased by 1.25 times.SGCSHI employs single-slit line-FOV imaging instead of multi-slit planeFOV imaging and turns scan detection into stare detection,where grating or prism serves as optical splitter to achieve data with high spectral resolution;the mean spectral resolution of the experiment system in this paper is as high as 3.8nm(@550nm).Furthermore,the slit movement based spectral imaging system(SSHI,Slit Scan Hyperspectral Imager)was proposed for detection with high spectral resolution;it offers the same detection efficiency as traditional detection methods.The information acquired per cycle by SSHI is the coding "atom" of information obtained with SGCSHI system,serving as the basis for SGCSHI spectral imaging.SGCSHI and SSHI offer a new route of methodology for dynamic target video detection to be realized through high-resolution hyperspectral imaging technology.2.The high-precision spectral recovery algorithm based on the joint of double sparse domains(JDSD)for seeking solutions was proposed against the low spectral fidelity of reconstruction results by traditional methods.The JDSD algorithm decomposes the signal into low-frequency signal and high-frequency signal,and performs different sparse decompositions depending on the characteristics of signal frequency distribution,thereby enabling the decoding to realize detail compensation;this overcomes the loss of details such as spectral absorption peak that is frequently observed in traditional reconstruction methods based on single sparse domain transform.In mass sample test at a sampling rate of 20%,the fidelity indicators SAM and GSAM of mean recovery result of the traditional reconstruction algorithm are 0.625 and 0.515,respectively,while the values of mean reconstruction results of the JDSD algorithm are 0.817 and 0.659,respectively;where the sampling rate is 80%,both indicators increase respectively from 0.863/0.808 to 0.940/0.897;the high-precision recovery result of JDSD is of great significance to the further application of the spectrum.3.The electrical slit group coded spectral imaging(E-SGCSHI,ElectricalSGCSHI)and the mechanical-movement multi-slit group coded(M-SGCSHI,Mechanical-movement SGCSHI)techniques were proposed,respectively.The ESGCSHI system employs liquid crystal light valve as electronic coding device to achieve high-speed transform coding and impressive control flexibility,but its poor optical utilization(less than 10%)brings about poor signal-to-noise ratios of the images obtained through the system.With glass mask plate as its coding device,the M-SGCSHI system works with the field diaphragm to achieve the code transformation through precise mechanical movement.With an optical utilization of over 90%,the glass mask plate featuring high contrast and other advantages remarkably improves the poor signalto-noise ratio of E-SGCSHI,having enabled the acquisition of spectral imaging information with high imaging quality;in the final outcome,the mean SNR is approx.31 d B for E-SGCSHI and over 40 d B for M-SGCSHI.4.The spectral fidelity assessment technique(GSAM,Gradient SAM)based on gradient information was proposed against the difficulty in objectively reflecting the fidelity of spectral fingerprint feature by traditional spectral similarity evaluation method.GSAM makes use of the first-order gradient information of spectral shape to enhance such fingerprint characteristics as spectral absorption peak,and brings it under the spectral fidelity assessment.GSAM is extremely sensitive to the changes of spectral characteristic peaks and other details;in the test data,the SAM is between 0.998 and 1 while the GSAM is between 0.72 and 1 as the sampling rate changes;the resulting greater discrimination in difference objectively reflects the system's fidelity capabilities for spectral characteristics.Furthermore,the GSAM is also applicable for terrain classification;according to the test data,the SAM-based mean classification accuracy is 0.86 while such accuracy of the GSAM-based approach is 0.93,which highlights the role of spectral features in classification.On the other hand,since GSAM uses gradient information as assessment criteria,it is not affected by offset,featuring higher robustness.