| 摘要: | 立方衛星體積小、重量輕、發射載具的成本低，使其發射需求逐年增加。鑒於立方衛星高靈活度的特性，適用於各種任務目標；我們開發了一個空間尺寸4U、重量1.4 kg的高光譜科學酬載，命名為HyperSCAN (Hyper SpecCral AnNlyzer)。此高光譜儀搭載在12U的鯨鯊號立方衛星，觀測的方式以推掃式進行地面資訊的線性掃描；將可見光以及近紅外光譜細分成162個波段，解析地面資料中的光譜特徵，提供地形地物分類以及輻射亮度量測的研究。高光譜儀搭載高感度CMOS感測器。此小型高光譜儀提供具有科學價值且低成本的影像資料，為未來搭載在衛星上的高光譜開發，進行先期的小型衛星測試實驗。
本論文工作開發具有科學價值資料，包含資料標準流程設計、科學化校正以及資料驗證三個方面。資料標準流程設計方面，分成3個資料層級，分別是Level 0衛星原始封包資料、Level 1高光譜立方資料 (包含空間以及光譜維度) 以及Level 2絕對輻射校正資料。資料流程設計可以幫助即時做後端的資料處理，提供地面資料標準化作業。高光譜資料校正方面，包含視野校正、波長校正以及絕對輻射校正。首先使用焦距無限遠的平行光進行視野校正，利用精密六軸進行旋轉角調整，測得視野寬度約 8.03度，對應500公里軌道，地面的像幅寬度70.2公里，沿軌長度33.7公里。第二，利用單光儀進行波長校正，測出波長範圍在可見-近紅外光波段430 - 800 nm，光譜解析度 < 10 nm，以550 nm的單色光來說 FWHM為8.72 nm。第三，感測器光訊號的輻射解析度為10 bits。因此，我們使用單光儀和積分球等光學儀器進行絕對輻射校正。其目的是將讀取到的DN值轉化成光譜輻射亮度 (watt/cm2-μm-sr)，感測器感測範圍會落在10-5~10-2 watt/cm2-μm-sr，符合高光譜所測量的大氣層頂輻射亮度範圍 (10-3~10-2 watt/cm2-μm-sr)。在資料訊噪比 (Signal-to-Noise Ratio, SNR) 的表現上，以單光儀資料計算SNR都有 > 15 dB。資料驗證方面，我們進行實際戶外側掃觀測，利用上述的資料流程以及資料校正後，我們與標準的輻射傳輸模型 (MODTRAN) 來做比較，量測的輻射亮度數量級為10-3~10-2 watt/cm2-μm-sr與MODTRAN所估計的大氣背景輻射亮度在同一個量級範圍。
另外我們分析目標物的絕對輻射亮度光譜特徵，包含草地、建築物、玻璃以及背景大氣。此外，也使用主成分分析進行資料降維，根據波段的數量得到162個特徵。依權重進行排序之後，以線性 SVM 進行高光譜資料的分類，若輸入有162個特徵的原始Level 2資料：分類準確度高達99%；當使用主成分分析降維至10個特徵：分類準確度一樣能維持99% 且訓練時間也降低。因此，適當的降維資料能減少高光譜儀資料的冗餘以及擷取光譜特徵，得到良好的分類效果。
本論文發展科學價值高光譜資料，所量測的輻射亮度符合在軌的需求，且經由實際觀測驗證，證明高光譜儀量測光譜特徵以及輻射亮度，能進行地面資料分類的應用，驗證其未來發展小型衛星高光譜儀的研究價值。;The small size and light weight of CubeSats, as well as the low cost of the launch vehicle, has driven an increase in progress in CubeSat missions. Due to the high flexibility of CubeSats, which are suitable for various mission objectives, we have developed a 4U space size, 1.4kg hyperspectral scientific payload named HyperSCAN (Hyper SpecCral AnNlyzer). The HyperSCAN is onboard the 12U SCintillation and IONsophere eXtended (SCION-X) CubeSat, using the pushbroom method to scan ground information linearly in sweeping mode. The visible and near-infrared spectra are subdivided into 162 bands, and the spectral features in the ground data are analyzed, which can be used to provide research on the classification of terrain and features as well as the measurement of absolute radiance. With a high-sensitivity CMOS sensor, HyperSCAN offers scientifically valuable and low-cost image data for the scientific community and develops the future hyperspectral imager onboard small satellite missions.
The aims of this thesis consist of three aspects: standard data processing flow, scientific calibration, and data validation. The data flow design is divided into three levels: Level 0 raw packet data, Level 1 hyperspectral cubic dataset (including two spatial and one spectral dimension), and Level 2 absolute radiometric calibration data. The data flow design can help the back-end data processing in real-time and provide ground data standardization.
The calibration process for HyperSCAN is a meticulous one, ensuring the accuracy and reliability of the data it provides. This includes field-of-view (FOV) correction, wavelength measurement, and absolute radiometric calibration. For FOV correction, we use parallel light with infinite focal length to carry out FOV experiment, and use precision six-axis to carry out rotation angle adjustment to measure the FOV of ~ 8.0 degrees, corresponding to the image width on the ground is 70.2 km at the 500 km orbit, and along-track length of 33.7 km. The wavelength correction was carried out by using a monochromator, and the wavelength was measured between 430 - 800 nm in the visible-near-infrared band. The FWHM was 8.72 nm for monochromatic light of 550 nm. We calibrated absolute radiometric calibration using a power meter and an integrating sphere for flat field measurement, and converted the readout 10-bit digital number (DN) value into the absolute spectral radiance (watt/cm2-μm-sr). The HyperSCAN ranges from 10-5 to 10-2 watt/cm2-μm-sr, within the typical range of the top-of-atmosphere radiance (10-3~10-2 watt/cm2-μm-sr). The SNR of HyperSCAN is > 15 dB.
We conducted an outdoor experiment for data validation and calibrated absolute spectral radiance using the above data processing flow. Compared with the standard radiation transmission model (MODTRAN), the measured radiance quantities range within the order of magnitude of 10-3~10-2 watt/cm2-μm-sr, closely matching the model results. Finally, we used the principal component analysis (PCA) to reduce the data dimensions where original data have 162 bands. After sorting the weights of the dataset, the linear Support Vector Machine was used to categorize the hyperspectral data. The classification accuracy was as high as 99% for a total of 162 bands. After PCA, if we downscaled the required spectral features to 10 dominant features, the classification accuracy was as high as 99% and required less training time, further demonstrating the reliability of our results. Therefore, the classification results demonstrate that the advantages of PCA, include data dimensionality reduction and spectral feature extraction.
This thesis develops HyperSCAN’s calibrations and data validation experiment, conforming to in-orbit requirements. We verified that the HyperSCAN performance could be applied to the ground data classification and satisfy the criteria of the future development of a small-satellite hyperspectral imager payload. |
