雲林地區因地下水使用密集與沖積層地質條件影響，長期面臨地層下陷問題。地層下陷具有時間累積性，且受地下水位變化、地層組成與深度條件等因素共同影響，因此如何利用既有監測資料建立具預測能力之模型，為地層下陷防治與地下水管理中重要之課題。
本研究結合地下水位測站、地層下陷監測井與鑽井岩性資料，建立不同深度之累積地層下陷量預測流程。研究首先以具梯度懲罰之提示式Wasserstein生成對抗插補網路（CWGAIN-GP）進行地下水位缺值補遺，並建立地下水位月變化量資料；地層下陷資料則依磁環深度位置推算不同深度之累積地層下陷量，地層下陷資料則依磁環深度位置推算不同深度之累積地層下陷量，並結合岩性分類與深度資訊建立模型資料集。其後，分別建立長短期記憶神經網路（LSTM）與極限梯度提升（XGBoost）模型，從時間方向與空間方向進行分析。其中LSTM以不同深度點之時間序列進行建模，XGBoost則將深度作為地層背景特徵直接輸入模型。時間方向用以探討模型對同一測站後續累積地層下陷量之預測能力；空間方向則利用鄰近測站資訊，推估未參與訓練測站之地層下陷變化。最後透過夏普利值加性解釋（SHAP）分析探討各輸入特徵對模型預測之影響。
研究結果顯示，CWGAIN-GP於地下水位補遺中可掌握整體水位變化趨勢，平均納許係數（NSE）達0.90，但在長時間連續缺失與高比例隨機缺失情境下，補遺不確定性仍明顯增加。時間方向預測結果顯示，LSTM與XGBoost皆能描述累積地層下陷量隨時間之變化，其中XGBoost對不同訓練長度與輸入變數組合之表現較為穩定；深度分層誤差分析則顯示，淺層地層之預測誤差相對較大。空間方向預測結果顯示，XGBoost於半年期多步遞推預測中之平均R²為0.881，高於以預測起始月觀測值固定延續之基準值，其平均R²為0.778；一年期結果雖受個別低變異測站影響而出現發散，但排除異常測站後，XGBoost多步遞推預測之平均R²仍可提升至0.835，優於基準值之0.706。SHAP分析結果顯示，前一月地下水位變化、前一月地層下陷變化與深度資訊為影響XGBoost預測結果之主要特徵，地下水位變化亦具有一定貢獻，而岩性分類之直接貢獻相對有限。整體而言，本研究所建立之流程可整合地下水位補遺、深度剖面下陷資料與機器學習模型，並說明歷史下陷資訊、地下水位變化與地層條件於累積地層下陷預測中之相對影響。

Yunlin County has long experienced land subsidence due to intensive groundwater extraction and the compressible characteristics of its alluvial deposits. This study developed an integrated machine-learning framework for predicting cumulative land subsidence using groundwater-level records, depth-dependent subsidence measurements, and lithological data. Missing groundwater-level records were reconstructed using a Cue Wasserstein Generative Adversarial Imputation Network with Gradient Penalty (CWGAIN-GP). Long Short-Term Memory (LSTM) and Extreme Gradient Boosting (XGBoost) models were employed for temporal and spatial prediction, and SHapley Additive exPlanations (SHAP) were used to interpret the XGBoost results. CWGAIN-GP achieved a mean Nash-Sutcliffe efficiency coefficient of 0.90; however, uncertainty increased as consecutive data gaps approached one year and when the proportion of randomly missing data exceeded 30%. For temporal prediction, both models captured cumulative subsidence trends. Under the best seven-year training configurations, LSTM and XGBoost configurations achieved mean R² values of 0.85 and 0.93, respectively, with XGBoost exhibiting greater stability when trained on only one year of data. For six-month spatial prediction, the mean rollout R² values were 0.881 for XGBoost and 0.780 for LSTM, compared with 0.778 for the persistence baseline. After the anomalous Jhennan well was excluded, the corresponding one-year rollout R² values were 0.835 and 0.776, respectively, compared with 0.706 for the baseline. SHAP analysis identified the previous month’s groundwater-level change, the previous month’s subsidence change, and depth as the most influential predictors, whereas lithological categories made only a limited direct contribution. Overall, the proposed framework integrates groundwater-level imputation, depth-dependent subsidence monitoring, and machine-learning models to predict cumulative land subsidence.

1. 賴典章、費立沅、江崇榮（2003）。臺灣地區地下水分區特性。收錄於水文地質調查與應用研討會論文集（頁1–24）。經濟部中央地質調查所。
2. 經濟部（2014）。地下水補注地質敏感區劃定計畫書：G0001濁水溪沖積扇。經濟部。
3. 經濟部水利署（2014–2021）。中華民國一〇三年至一一〇年臺灣水文年報。經濟部水利署。
4. Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein generative adversarial networks. International conference on machine learning, 214-223.
5. Bagheri-Gavkosh, M., Hosseini, S. M., Ataie-Ashtiani, B., Sohani, Y., Ebrahimian, H., Morovat, F., & Ashrafi, S. (2021). Land subsidence: A global challenge. Science of the Total Environment, 778, 146193.
6. Chaussard, E., Wdowinski, S., Cabral-Cano, E., & Amelung, F. (2014). Land subsidence in central Mexico detected by ALOS InSAR time-series. Remote sensing of environment, 140, 94–106.
7. Chen, B., Gong, H., Chen, Y., Li, X., Zhou, C., Lei, K., Zhu, L., Duan, L., & Zhao, X. (2020). Land subsidence and its relation with groundwater aquifers in Beijing Plain of China. Science of the Total Environment, 735, 139111.
8. Chen, H.-Y., Vojinovic, Z., Lo, W., & Lee, J.-W. (2023). Groundwater level prediction with deep learning methods. Water, 15(17), 3118.
9. Chen, T., & Guestrin, C. (2016). Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 785-794.
10. Eghrari, Z., Delavar, M., Zare, M., Beitollahi, A., & Nazari, B. (2023). Land subsidence susceptibility mapping using machine learning algorithms. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 10, 129–136.
11. Galloway, D. L., Hudnut, K. W., Ingebritsen, S., Phillips, S. P., Peltzer, G., Rogez, F., & Rosen, P. (1998). Detection of aquifer system compaction and land subsidence using interferometric synthetic aperture radar, Antelope Valley, Mojave Desert, California. Water resources research, 34(10), 2573–2585.
12. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735–1780.
13. Hosseinzadeh, E., Anamaghi, S., Behboudian, M., & Kalantari, Z. (2024). Evaluating machine Learning-Based approaches in land subsidence susceptibility mapping. Land, 13(3), 322.
14. Kratzert, F., Klotz, D., Brenner, C., Schulz, K., & Herrnegger, M. (2018). Rainfall–runoff modelling using long short-term memory (LSTM) networks. Hydrology and Earth System Sciences, 22(11), 6005–6022.
15. Li, F., Liu, G., Tao, Q., & Zhai, M. (2023). Land subsidence prediction model based on its influencing factors and machine learning methods. Natural hazards, 116(3), 3015–3041.
16. Liu, J., Liu, W., Allechy, F. B., Zheng, Z., Liu, R., & Kouadio, K. L. (2024). Machine learning-based techniques for land subsidence simulation in an urban area. Journal of Environmental Management, 352, 120078.
17. Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in neural information processing systems, 30.
18. Rahmati, O., Falah, F., Naghibi, S. A., Biggs, T., Soltani, M., Deo, R. C., Cerda, A., Mohammadi, F., & Bui, D. T. (2019). Land subsidence modelling using tree-based machine learning algorithms. Science of the Total Environment, 672, 239–252.
19. Rahmati, O., Golkarian, A., Biggs, T., Keesstra, S., Mohammadi, F., & Daliakopoulos, I. N. (2019). Land subsidence hazard modeling: Machine learning to identify predictors and the role of human activities. Journal of Environmental Management, 236, 466–480.
20. Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592.
21. Shi, L., Gong, H., Chen, B., & Zhou, C. (2020). Land subsidence prediction induced by multiple factors using machine learning method. Remote Sensing, 12(24), 4044.
22. Smith, R. G., & Majumdar, S. (2020). Groundwater storage loss associated with land subsidence in Western United States mapped using machine learning. Water resources research, 56(7), e2019WR026621.
23. Wang, Y., Xu, X., Hu, L., Fan, J., & Han, M. (2024). A time series continuous missing values imputation method based on generative adversarial networks. Knowledge-Based Systems, 283, 111215.
24. Yazbeck, J., & Rundle, J. B. (2023). Predicting short-term deformation in the central valley using machine learning. Remote Sensing, 15(2), 449.
25. Yoon, J., Jordon, J., & Schaar, M. (2018). Gain: Missing data imputation using generative adversarial nets. International conference on machine learning, 5689-5698.
26. Zhou, C., Gong, H., Chen, B., Li, X., Li, J., Wang, X., Gao, M., Si, Y., Guo, L., & Shi, M. (2019). Quantifying the contribution of multiple factors to land subsidence in the Beijing Plain, China with machine learning technology. Geomorphology, 335, 48–61.