以衛星影像與光譜混合分析方法監測大阿里山茶區之崩塌地

Abstract

臺灣山區的茶產業是重要的經濟活動，但過去研究指出，森林轉為茶園等土地使用變化，可能提升崩塌與土壤流失風險。然而，茶樹具有密集根系且多配備完善排水系統，也可能發揮穩定坡地的正面效果。本研究以大阿里山地區為對象，結合中高解析度衛星影像與空間分析技術，深入探討茶園是否對崩塌地發生具有顯著影響。首先，我們以U-net深度學習模型結合Sentinel-2多時序影像、數值地表模型與地面實測資料，進行土地使用分類，獲得高準確度（整體準確度與茶園的使用者/生產者準確度均超過90%）的分類結果。接著，使用Landsat地表反射率影像進行Bayesian MAP光譜校正，確保多源影像的一致性，再以AutoMCU光譜混合分析法（SMA）推估崩塌地之土壤比例，建構跨年度的崩塌地分布圖。最後，透過環域分析（buffer analysis），計算不同土地使用類型周圍各距離範圍內之崩塌地面積與變化量。結果顯示，整體崩塌地面積在近20年間有下降趨勢。環域分析亦發現，茶園周圍400公尺內的崩塌風險相對較高，但其變化幅度不若森林明顯，且土壤比例變化速率在400公尺後趨於平緩，顯示其影響具有空間限制。整體而言，茶園並未導致明顯高於其他地類的崩塌風險，反而可能因管理設施而具有緩坡穩定的潛力。該結果有助於重新評估山地農業與地景穩定性之間的關係，並作為未來永續農業規劃的參考依據。

Tea cultivation is a key component of mountainous agriculture in Taiwan. While previous studies have suggested that land use changes—particularly forest conversion to tea plantations—may increase the risk of landslides and soil erosion, tea plantations also feature dense root systems and well-designed drainage infrastructure, potentially mitigating such hazards. This study investigates whether tea plantations contribute to landslide occurrence in the Greater Ali Mountain region by integrating remote sensing, spectral analysis, and spatial modeling. We first applied a U-net deep learning model to classify land use and land cover (LULC) using multi-season Sentinel-2 imagery, a digital surface model (DSM), and ground truth data. The resulting LULC maps achieved high classification performance, with overall accuracy and both user’s and producer’s accuracy for tea plantations exceeding 90%. To generate consistent surface reflectance values across Landsat imagery, we conducted Bayesian Maximum A Posteriori (MAP) spectral correction. Landslide mapping was then performed using the AutoMCU spectral mixture analysis (SMA) model, which estimates sub-pixel fractions of photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), and soil, enabling accurate landslide delineation based on elevated soil fractions. Buffer analysis was employed to assess the spatial distribution of landslide-prone areas around different LULC types. Results show that total landslide area has decreased steadily over the past two decades. While a higher landslide density was observed within 400 meters of tea plantations, the magnitude of change was less pronounced than that surrounding forests. The rate of change in soil proportion also declined beyond the 400-meter threshold, indicating spatially limited influence. These findings suggest that tea plantations do not significantly exacerbate landslide hazards and may, in some cases, contribute to slope stability. This research offers new insights into the complex relationship between mountainous agriculture and landscape hazards and provides a valuable reference for sustainable land management in hilly regions.