基于可解释性深度学习的星云湖表层水温预测研究

Prediction of surface water temperature in Xingyun Lake based on interpretable deep learning

摘要

【目的】全球变暖背景下,湖泊表层水温(Lake Surface Water Temperature,LSWT)的持续上升已成为加剧湖泊富营养化进程的关键因素,因此准确预测其未来变化趋势对湖泊富营养化预警及水环境管理具有重要意义。【方法】基于MODIS遥感影像获取2002—2024年星云湖逐日表层水温数据,采用随机森林(Random Forest,RF)算法插补其缺失值,进而构建并行的LSTM-Transformer深度学习模型预测未来不同气候情景(SSPs)下LSWT的变化趋势,并引入夏普利加性解释法(SHapley Additive exPlanations,SHAP)揭示LSWT的关键驱动因子及其影响方向。【结果】结果表明:(1)2002—2024年期间星云湖表层水温的升温速率为0.27℃/10 a,秋季升温趋势尤为突出。(2)LSTM-Transformer融合模型具有较高的预测精度,决定系数(R²)为0.95,均方根误差(RMSE)为0.89℃;在未来气候情景(SSP1-2.6、SSP2-4.5和SSP5-8.5)下,星云湖表层水温预计将以0.11~0.42℃/10 a的速率持续上升。(3)近地表气温和地表下行长波辐射是影响星云湖表层水温变化的关键驱动因子。【结论】本研究构建的深度学习模型在高原湖泊表层水温预测中具有良好的适用性,该框架可为高原湖泊在未来气候情景下的水环境适应性管理与生态风险预警提供依据。

Abstract

[Objective] In the context of global warming, the continuous rise in lake surface water temperature(LSWT) has become a key factor in exacerbating lake eutrophication. Therefore, accurately predicting its future trends is of great significance for the early warning of lake eutrophication and water environment management. [Methods] Based on MODIS remote sensing images, daily LSWT data for Xingyun Lake from 2002 to 2024 were acquired. The random forest(RF) algorithm was employed to impute missing values. Subsequently, a parallel LSTM-Transformer deep learning model was developed to predict future LSWT trends under different climate scenarios(SSPs). Furthermore, the SHapley Additive exPlanations(SHAP) method was introduced to identify the key driving factors of LSWT and their directions of influence. [Results] The results indicated that:(1) the warming rate of surface water temperature in Xingyun Lake was 0. 27 ℃/10 a from 2002 to 2024, with a particularly pronounced warming trend in autumn.(2) The LSTM-Transformer fusion model exhibited high prediction accuracy, with a coefficient of determination(R²) of 0. 95 and a root mean square error(RMSE) of 0. 89 ℃. Under future climate scenarios(SSP1-2. 6, SSP2-4. 5, and SSP5-8. 5), the LSWT of Xingyun Lake was projected to continue rising at a rate of 0. 11~0. 42 ℃/10 a.(3) Near-surface air temperature and surface downward longwave radiation were identified as the primary driving factors influencing LSWT variations in Xingyun Lake. [Conclusion] The developed deep learning model exhibits good applicability for predicting surface water temperature in plateau lakes. The framework can provide a basis for adaptive water environment management and early warning of ecological risks in plateau lakes under future climate scenarios.

关键词

湖泊表层水温 / MODIS / 星云湖 / 深度学习模型 / LSTM-Transformer / 气候情景 / SHAP / 驱动因子分析

Key words

lake surface water temperature / MODIS / Xingyun Lake / deep learning model / LSTM-Transformer / climate scenarios / SHAP / driving factor analysis

基金资助

国家重点研发计划(2023YFC3207503)