| 摘要: | 本論文探討使用微波輻射量測地表參數，發展一個具相關誤差之倒傳遞神經網路，處理輻射亮溫與植被和土壤含水量的非線性映射。該神經網路模擬實際的生理現象，例如皮層上的某一點受到刺激，該點內具有最多神經末梢的神經纖維受到刺激，鄰近的神經纖維接受到較輕微的刺激，愈遠的纖維受到的刺激愈弱，即可建構為一個具相關誤差之倒傳遞神經網路。隨後將該神經網路應用在遙測領域----從亮溫反演植被和土壤含水量。土壤水分支配了地表與大氣間輻射能如何分割成感熱及潛熱。因此，在水文、氣象、農業和生物地球化學的理論上，扮演極為重要的角色，成為在遙測領域中很感興趣的參數。 本論文的具體成果可分述如下： 1. 具相關誤差之倒傳遞神經網路的發展 具相關誤差之倒傳遞神經網路(A Backpropagation Neural Network with Error Correlation)依照輸出層神經元間距離遠近定出不同的相關性，用來描述鄰近神經元的交互作用。實驗証明適當的相關誤差能提高神經網路訓練的精確度與準確度。具相關誤差之倒傳遞神經網路具有二個調整參數，首先距離參數α是用來調整相關誤差的的有效距離，其次強度參數λ是用來調整相關誤差的強度。在距離α外的神經元當作是無關的。 2. 應用在遙測領域 ~ 從亮溫反演土壤含水量:以模式模擬結果為例 在這個研究中，使用相關誤差之倒傳遞神經網路，從模擬亮溫反演土壤表面的含水量。使用的頻率包括先進微波掃瞄輻射計AMSR (Advanced Microwave Scanning Radiometer) 之6.9、10.7 GHZ 和土壤水分海洋鹽度感應器SMOS (Soil Moisture and Ocean Salinity sensor) 之1.4 GHZ。本研究使用的模式為LSP/R (Land Surface Process/ Radiobrightness) ， 該模式提供一系列的土壤水分和亮溫資料。AMSR 的觀察角度為55度，而SMOS的觀察角度則介於0到55度間，為了方便本研究，吾人使用了0、10、20、30、40、50度等多重觀察角度。這些多重的頻率和角度，允許吾人設計不同的觀測方式來檢驗它們對於土壤含水量的敏感性。例如; L-band 單一觀看角度視為L-band 1-D觀測方式。同時，它可組合其他觀看角度，變成L-band 2-D 或多維的觀測方式，又或者結合AMSR 6.9、10.7 GHZ，變成多頻率/維度的觀測方式。本研究中，顯示L-band 1-D觀測方式亮溫對於土壤含水量是敏感的，且此敏感性可以用整合L-band第二觀察角度或AMSR 頻道來增加。 3. 應用在遙測領域 ~ 從亮溫反演植被和土壤含水量:以田野資料為例 在這個研究中使用法國PORTOS(PORTOS-93，96 experiment)輻射計在小麥田小麥生長期連續三個月收集的田野資料。從PORTOS-93資料中隨機選取作為神經網路訓練和測試，除了從亮溫反演土壤含水量並進一步反演植被含水量，我們發現利用L-band 2-D觀測方式時，可得到最好的植被和土壤含水量。此外，訓練過的神經網路以PORTOS-96資料作進一步的評估。研究顯示，反演土壤含水量的平均誤差率約為4%，植被總含水量平均誤差率約0.239kg/m2 (0.160kg/m2，93年和0.319kg/ m2，96年) ，雖然只使用1993年資料於網路訓練，因1996年的反演結果令人相當滿意，使我們對於使用神經網路從亮溫反演植被和土壤含水量充滿信心。 This dissertation is intended to investigate the sensing of surface parameters by microwave radiometry. A backpropagation neural network with error correlation (BNNEC) is developed to manage the nonlinear relationship between surface parameters and radiometric signatures. Its performance of retrieving plant water content (PWC) and soil moisture content (SMC) from brightness temperatures is examined by using both predictions from model simulations and measurements from field experiments. The backpropagation neural network with error correlation (BNNEC) incorporates a novel rule, error correlation learning, to train a feedforward neural network. The correlated error terms associated with nearby neurons differ from those of the existing neural networks without considering of the correlations. That is, the output layers of its neurons are trained simultaneously and interactively through the error correlation terms. The BNNEC is applied to retrieve soil moisture from simulated brightness temperatures and perform the iris classification. Simulation results demonstrate the superiority of the proposed BNNEC. In addition, we optimize the observing scheme for sensing surface soil moisture (SM) from simulated brightness temperatures by the BNNEC. The frequencies of interest include 6.9 and 10.7 GHz of the Advanced Microwave Scanning Radiometer (AMSR), and 1.4 GHz (L-band) of the Soil Moisture and Ocean Salinity (SMOS) sensor. The Land Surface Process/Radiobrightness (LSP/R) model is used to provide time series of both SM and brightness temperatures at 6.9 and 10.7 GHz for AMSR's viewing angle of 55 degrees, and at L-band for SMOS's multiple viewing angles of 0, 10, 20, 30, 40, and 50 degrees. These multiple frequencies and viewing angles allow us to design a variety of observation modes to examine their sensitivity to SM. For example, L-band brightness temperature at any single look angle is regarded as an L-band 1D observation mode. Meanwhile, it can be combined with either the observation at other angles to become an L-band 2D mode or a multiple dimensional observation mode, or with the observation at 6.9 or 10.7 GHz to become a multiple frequency/dimensional observation mode. In this study, it is shown that the L-band 1D radiometric observation is sensitive to SM. The sensitivity can be increased by incorporating radiometric observation either from a second angle, or from multiple look angles, or from any of the two lowest AMSR channels. In addition, the advantage of an L-band 2D mode or a multiple dimensional observation mode over an L-band 1D observation mode is demonstrated. Moreover, we investigate the best observing configuration for sensing wheat plant water content (PWC) and soil moisture content (SMC) profiles from the measured H- and V-polarized brightness temperatures at 1.4 (L-band), and 10.65 (X-band) GHz by the BNNEC. The brightness temperatures were taken by the PORTOS radiometer over wheat fields through 3 months growth cycles in 1993 (PORTOS-93) and 1996 (PORTOS-96). During both field campaigns, the radiometer was used to measure brightness temperatures at incident angles from 0 to 50 degrees at L-band and at an incident angle of 50 degrees at X-band. The SMC profiles were measured to the depths of 10 cm in 1993 and 5 cm in 1996. The wheat was sampled approximately once a week in 1993 and 1996 to obtain its dry and wet biomass (i.e., PWC). The BNNEC was trained with observations randomly chosen from the PORTOS-93 data, and evaluated by the remaining data from the same set. The trained neural network is further evaluated with data from the PORTOS-96. |
