| 摘要: | 土壤表層濕度資訊對於水資源管理非常重要，同時對於農作物管理與產量推估 上也有賴於作物生長區域的資訊。本研究旨在應用中尺度影像光譜儀(MODIS)衛星影像 來探討越南湄公河三角洲土壤表層濕度變化與稻米作物耕種型態的相關性。土壤表層 濕度主要是根據2002 年至2007 年，每年一月到四月的MODIS 影像資料推導估計溫度 植生乾旱指標(TVDI)而來；此指標主要是利用MODIS 影像資料的地表溫度(LST)與植生 差異指標以經驗參數化推導而來。 從土壤表層濕度的估計結果可知，2006 年研究區的低土壤表層濕度區域範圍相 較於其他研究年分而言是最大的，因此本研究選擇以2006 為極端年與2002 為平常年， 分析這兩年的土壤表層濕度變化與稻米耕作型態分布的關聯性。稻作型態的空間分布 主要是以2002 年與2006 年MODIS NDVI 250 公尺解析度的時序影像資料進行分類後產 製：MODIS NDVI 時序資料先以經驗模態分解法(EMD)過濾雜訊；其後，再分別以硬分類： 線性混合模式(LMM) ，及軟分類：支持向量機(SVMs)分類演算法分別進行稻作型態之 分類及製圖。這兩種分類演算方法使用的目的是比較它們的分類優劣，同時本研究也 使用不同的空間與非空間資料來評估TVDI與稻作分類的精度。 推導結果顯示LST-NDVI 像點資料的空間分佈呈現非常清楚的三角形散佈特徵： 這表示研究區土壤表層濕度有大範圍的分布差異。其後再藉由逐日降雨資料來驗證研 究區的TDVI。結果顯示TDVI 與逐日降雨資料有相當的一致性與敏感性。從2002 年到 2005 年，較低的土壤表層濕度區域主要分布在海岸地區；但在2006 與20007 年就逐漸 擴展到三角洲中部區域。同時2006 年，低土壤表層濕度的區域範圍達到最大，呼應著 在2006 年當時湄公河三角洲區域，湄公河與巴塞河在旱季時因特殊原因所致逕流量大 減所造成的乾旱現象。 vii 在稻米生長週期偵測案例研究中，以過濾剖線偵測出的播種與高株分蘗期的日 期與田野調查資料的比較結果中顯示：使用EMD 濾波相較於小波(wavelet)濾波方法， EMD 過濾的稻作生長曲線保持較佳的稻作生長NDVI 資訊。這些NDVI 濾波的剖線型態反 映了不同稻作耕種型態在不同稻米成長季節的變化。通過了解不同稻作型態的NDVI 時 序剖線特徵，接著以軟分類LMM 與硬分類SVMs 的分類方法分類由EMD 濾波的NDVI 時 序剖線，以產製研究區的稻作型態分類。然後2002 年與2006 年稻作分類圖再以地真 資料與政府統計資料進行比對。比對結果顯示，對研究區的稻作分類而言，LMM 與 SVMs的兩種分類法都是極佳的分類方法。 由分類結果與地真資料的比較可知，SVMs 的分類成果較LMM 的分類成果稍佳。 SVMs的總體分類精度與Kappa 係數在 2002 年分別為84.0% 與 0.79；2006 年為85.1% 與0.80；LMM 的分類成果值較SVMs 低；其總體分類精度與Kappa 係數在 2002 年分別 為81.8%與0.76,；2006 年分別為79.9%與0.73。同時這些由MODIS 資料推導出的稻作 分類結果皆與官方省級尺度的稻作統計資料有很高相關性(R2 > 0.85)。兩種分類方法 以z-test 檢測分類差異，在2002 年與2006 年的估計值分別為 0.299 與 0.275，皆 小於95%的信賴區間值1.96，通過檢測，顯示兩種分類方法並這兩年皆無統計上的分 類差異。 對於土壤表層濕度與稻作形態的關聯性分析，本研究以2002 年與2006 年土壤 表層濕度較 乾與非常乾類別的組圖與稻米分類區進行比較。結果顯示在該時期，土壤 表層濕度較乾與非常乾的地區， Information on surface soil moisture is important for water management, while information on rice growing areas is vital for crop management and production prediction. This study aims to investigate surface soil moisture variability in relation to rice cropping systems in the Mekong Delta (MD), Vietnam using the Moderate Resolution Imaging Spectroradiometer (MODIS) data. The surface soil moisture was estimated from the MODIS data acquired during January to April from 2002 to 2007 using the Temperature Vegetation Dryness Index (TVDI) method. This index was empirically calculated by parameterizing the relationship between the MODIS Land Surface Temperature (LST) and the Normalized Difference Vegetation Index (NDVI) data. From the results of soil moisture estimation, it was found that the low soil moisture occurred in 2006 and occupied the largest region of the study area compared to other years. Therefore, this extreme year 2006 and a normal year, in this case 2002, were selected for analysis of soil moisture variability in relation to the distribution of rice cropping systems. The spatial distribution of rice cropping systems was obtained from classification of the timeseries MODIS NDVI 250-m data acquired in 2002 and 2006. Data were processed using the empirical mode decomposition (EMD) method for noise filtering of the time-series NDVI data. Soft and hard classification algorithms, namely linear mixture model (LMM) and support vector machines (SVMs), were used for classifying rice cropping systems. These two classification algorithms were used for the sake of comparing their classification performance. Various spatial and non-spatial data were also gathered for accuracy assessment of the TVDI and classification results. The results showed that the LST-NDVI space was well-defined. The pixels in each scatter plot could form a triangle. This indicated a wide range of surface soil moisture in the study area. The TVDI validation results were achieved by comparing TVDI values with daily ix rainfall throughout the study area. The comparison results revealed good agreement and sensitivity between TVDI and daily rainfall data. The areas with low soil moisture were mainly distributed in coastal areas from 2002 to 2005, but expanded into the middle region in 2006 and 2007. The largest area of low soil moisture was observed in 2006, reflecting the fact that the MD was faced with drought in 2006 because the amount of water in the Mekong and Bassac Rivers in the dry season was reduced drastically. In a case study of rice crop phenology detection, the comparison results between the estimated sowing/heading dates and the field survey data indicated that the use of smooth time profiles extracted from the EMD-based filtered time-series MODIS NDVI 250-m data for detecting phenological dates gave better results than the wavelet transform-based data. The EMD acted a good filter for noise reduction of the time-series NDVI data. The smooth NDVI profiles extracted from the EMD-based filtered NDVI data could well preserve the amplitude of NDVI values better than those extracted from the wavelet transform. These NDVI patterns reflected the seasonal changes in crop phenology of rice cropping systems, which was important for understanding the temporal NDVI responses of different rice fields of cropping patterns in the study area. The LMM and SVMs were applied to the EMD-based filtered data for classification of rice cropping systems in the region. The classification maps for 2002 and 2006 were compared with the ground truth data and government rice area statistics. The comparison results indicated that both classification methods (LMM and SVMs) were promising for rice crop mapping in the region. The comparison results between the classification results and the ground truth data indicated that the SVMs gave slightly better classification results than the LMM. The overall accuracy and Kappa coefficient achieved by the SVMs for the year 2002 data were 84.0% and 0.79, while the values for the LMM were 81.8% and 0.76, respectively. Similarly, the overall accuracy and Kappa coefficient achieved by the SVMs for the year 2006 data were 85.1% and x 0.80, and those for the LMM were 81.8% and 0.76, respectively. These comparison results reaffirmed good agreement between the MODIS-derived rice areas with the government rice area statistics at the provincial level (R2 > 0.85 in all cases). However, a significance test of difference between two classification methods using Z-test method revealed that the classification accuracy between these two classification methods (i.e., LMM and SVMs) were not statistically significant different. The Z-test values between the classification methods reported for the year 2002 and 2006 data were 0.299 and 0.275, respectively. These values were smaller than the critical value of 1.96. To relate surface soil moisture variations with rice cropping systems, the composite soil moisture maps (considering dry and very dry classes) were aggregated with the rice crop maps for the years 2002 and 2006. The results indicated a remarkable increase in the area of double and triple irrigated rice cropping systems in areas of low soil moisture (i.e., dry and very dry conditions) during this period. Approximately, 6.3% and 9.9% of the area of double and triple irrigated rice cropping systems identified as low soil moisture in 2002 increased to 14.9% and 16.3% in 2006, respectively. This study has demonstrated merits of using MODIS data for studying soil moisture variability in relation to rice cropping systems, which is important for crop and water management. |
