六龜人工林與崩塌地干擾對森林長期動態影響之研究

Abstract

臺灣森林因地處颱風頻繁且地質脆弱之地區，長期面臨天然災害與人工林退化的雙重壓力，導致森林結構與功能產生劇烈變動。為掌握森林演替與災後復育之動態，本研究以六龜試驗林為研究區，結合多時期衛星影像、空載光達與地理資訊系統，運用邏輯迴歸 (Logistic Regression, LG)、隨機森林 (Random forest, RF) 與細胞自動機-馬可夫鏈 (Cellular automaton and Markov Chain, CA-Markov) 等模型，進行森林變遷與崩塌復育潛勢之時空分析與預測。本研究首先探討臺灣杉人工林的自然演替變化與地理因子關聯，並建立全區臺灣杉人工林變化之邏輯迴歸與隨機森林預測模型。其後考量造林地區位差異，分別建立北鳯岡林區與南多納林區的分區模型，結果顯示分區模型能顯著提升預測準確性。其中RF模型準確率高達94.8%與97.8%，AUC值分別為0.988與0.998，研究成果顯示高程、雨量與距林道距離為主要影響因子。在崩塌與復育長期監測方面，莫拉克颱風後六龜崩塌整體呈現劇烈減少、後趨穩定變化趨勢，崩塌地景破碎化指標，也呈現下降趨勢，且地景連續性增加，顯示結構穩定性提升。崩塌復育則發現在崩塌發生後初期階段復育速度較快，較後期更具恢復力。在模型建構方面，空間解析度經測試以10 m解析度表現較佳。在模型方法上，以隨機森林較邏輯迴歸表現較好。其中崩塌潛勢RF模型，分別在2009 - 2017與2017 - 2024兩個時段建置模型，其AUC值皆超過0.950，準確度分別為89%及90%，模型表現極佳，主要影響因子為冠層結構指標、NDVI與高程。崩塌復育RF模型AUC值則分別為0.87與0.90，準確度為77%與83%，表現亦佳，關鍵因子為距崩塌邊緣距離、高程及坡度，顯示地形與種源供應為復育重要因素。在未來土地覆蓋變化預測部份，首先利用2009年與2017年兩期土地覆蓋型變化資料及潛在影響因子，結合RF模型產生轉移機率圖，並整合CA-Markov模型，模擬預測2024年之土地覆蓋空間分布。預測結果與實際2024年土地覆蓋型進行比對，並與傳統僅使用CA-Markov模型的預測效能進行比較。結果顯示RF-CA-Markov模型準確率達95.78%，Kappa值0.522，顯著優於傳統CA-Markov模型之準確率 (84.88%) 與Kappa值 (0.253)。完成模型驗證後，進一步以2017年至2024年期間之土地覆蓋變化與驅動因子進行RF模型訓練，重新產生轉移機率圖，並模擬2031年土地覆蓋變遷。模擬結果顯示，植生及崩塌地面積與2024年預測結果相近，未持續擴大，森林覆蓋趨於穩定，反映出災後崩塌與復育趨於穩定。 本研究證實整合多時期遙測資料與機器學習模型，可有效掌握森林演替與災害復育動態，提升崩塌潛勢與森林變遷預測能力。研究成果可提供森林經營與防災調適政策制定的量化依據，以促進森林資源之永續發展與生態系功能之長期穩定。

Taiwan’s forest ecosystems, residing in a landscape highly susceptible to typhoons and geological instability, face the dual pressures of natural disasters and plantations degradation, leading to significant changes in forest structure and function. To understand the dynamics of forest succession and post-disaster recovery, this study focused on the Lioukuei Experimental Forest, integrating multi-temporal satellite imagery, airborne LiDAR, and Geographic Information Systems (GIS). Logistic regression, Random Forest (RF), and Cellular Automaton-Markov Chain (CA-Markov) models were applied to conduct spatiotemporal analysis and prediction of forest changes and landslide recovery potential.This study first investigates the natural succession dynamics of Taiwania cryptomerioides plantations and their relationship with geographic factors, establishing logistic regression and random forest models to predict the spatial distribution of forest transitions across the entire study area. To account for spatial heterogeneity between afforestation sites, separate zonal models were developed for the northern Fenggang and southern Duona forest districts. Results indicated that zonal models significantly improved prediction accuracy, with RF models accuracies of 94.8% and 97.8%, and AUC values of 0.988 and 0.998, respectively. Key influencing factors identified were elevation, precipitation, and distance to forest roads.In terms of long-term landslide and vegetation recovery monitoring, post-Typhoon Morakot landslide activity in the Lioukuei region showed a marked initial decline followed by stabilization. Landscape fragmentation indices also exhibited decreasing trends, while landscape connectivity increased, indicating improved structural stability. Vegetation recovery was observed to proceed more rapidly in the early stages following landslides, suggesting higher resilience during the initial recovery phase. Model development showed that a spatial resolution of 10 m provided the best performance. Among modeling approaches, RF outperformed logistic regression. For landslide susceptibility, RF models were developed for the 2009–2017 and 2017–2024 periods, AUC values above 0.950 and accuracies of 89% and 90%, respectively. Key variables included canopy structure indices, NDVI, and elevation. For vegetation recovery modeling, RF models achieved AUC values of 0.87 and 0.90, with accuracies of 77% and 83%. The most influential variables were distance to landslide edge, elevation, and slope, indicating that terrain characteristics and seed source proximity are critical to recovery.For future land cover change prediction, RF models were first trained using land cover transitions and driving factors from 2009 to 2017 to generate transition probability maps, which were integrated with the CA-Markov model to simulate land cover distribution in 2024. Comparison with actual 2024 land cover data and the traditional CA-Markov model revealed that the RF-CA-Markov model outperformed the latter, achieving an overall accuracy of 95.78% and a Kappa coefficient of 0.522, significantly higher than the CA-Markov model’s 84.88% accuracy and 0.253 Kappa value.Following model validation, land cover transitions and driving factors from 2017 to 2024 were used to retrain the RF model, enabling simulation of land cover changes in 2031. Results showed that forest and landslide areas in 2031 remained similar to the 2024 projections, indicating no significant expansion of landslide zones and suggesting that post-disturbance vegetation recovery and landscape stabilization had reached a relatively steady state.This study confirms that integrating multi-temporal remote sensing data with machine learning models can effectively capture forest succession and disaster recovery dynamics, enhancing the predictive capability for landslide susceptibility and forest changes. The findings provide a quantitative basis for forest management and disaster adaptation policy, promoting the sustainable development of forest resources and the long-term stability of ecosystem functions.