Exploring stable isotope patterns in monthly precipitation across Southeast Asia using contemporary deep learning models and SHapley Additive exPlanations (SHAP) techniques

Stable isotopes are crucial for understanding water cycles and climate dynamics, particularly in tropical regions. However, establishing and maintaining precipitation sampling stations in Southeast Asia is challenging due to high costs and logistical issues. Consequently, many areas in this region have limited or no sampling stations with adequate stable isotope data. To address this problem, developing models that simulate stable isotope contents using machine learning (ML) techniques, especially deep learning, is a promising solution. In this study, the influence of large-scale climate modes (teleconnection indices) and local meteorological parameters on the stable isotope contents of precipitation was examined across six key stations in Southeast Asia, including Bangkok, Kuala Lumpur, Jakarta, Kota Bharu, Jayapura, and Singapore. A deep neural network (DNN) model was applied for simulation, and its performance was compared with a partial least squares regression (PLSR) model using various evaluation metrics. The DNN consistently demonstrated superior accuracy across all studied stations, highlighting the efficacy of DNNs, in accurately simulating stable isotope contents in tropical precipitation. The importance ranking derived from the SHapley Additive exPlanations (SHAP) technique aligns perfectly with the results obtained from the DNN importance function. In addition, the SHAP summary plot highlights the contributions of key features, such as precipitation and potential evaporation, to the model's predictions. The dependence plots further illustrate the relationship between these features and the predicted response, revealing nonlinear interactions that influence model behaviour. This research provides new insights into the complex interactions between large-scale climate drivers and local weather patterns, advancing the use of ML for isotope-based climate studies. The techniques used in this study offer a framework for applying ML to isotope analysis in tropical climates and can be extended to similar regions worldwide.

稳定同位素对于理解水循环和气候动力学至关重要，尤其在热带地区。然而，由于成本高昂和后勤问题，在东南亚建立和维护降水采样站具有挑战性。因此，该区域许多地区的采样站数量有限，或缺乏拥有充足稳定同位素数据的站点。为解决这一问题，利用机器学习（ML）技术（尤其是深度学习）开发模拟稳定同位素含量的模型是一种有前景的解决方案。本研究在东南亚六个关键站点（包括曼谷、吉隆坡、雅加达、哥打巴鲁、查亚普拉和新加坡），考察了大尺度气候模式（遥相关指数）和本地气象参数对降水稳定同位素含量的影响。研究应用深度神经网络（DNN）模型进行模拟，并通过多种评估指标将其性能与偏最小二乘回归（PLSR）模型进行比较。深度神经网络在所有研究站点均持续展现出更优的准确性，凸显了其在精准模拟热带降水稳定同位素含量方面的有效性。通过沙普利加性解释（SHAP）技术得到的重要性排序与深度神经网络重要性函数的结果完全一致。此外，SHAP摘要图凸显了降水和潜在蒸发等关键特征对模型预测的贡献；依赖图进一步阐明了这些特征与预测响应之间的关系，揭示了影响模型行为的非线性交互作用。本研究为大尺度气候驱动因子与本地天气模式之间的复杂交互作用提供了新的见解，推动了机器学习在基于同位素的气候研究中的应用。本研究采用的技术为机器学习在热带气候同位素分析中的应用提供了框架，并可扩展至全球其他类似地区。