作者：Li, T., Wang, Y., Wang, L. et al.

The Tibetan Plateau (TP), widely known as the "Third Pole" and "Asian Water Tower", hosts lakes that act as crucial indicators of climate change. While most TP lakes have expanded significantly in recent decades, gaps in high-temporal-resolution data and unclear scale-dependent driving mechanisms remain. Focusing on inner TP (ITP) lakes during 1972-2021, we used the self-attention-based imputation for time series (SAITS) deep-learning model to fill remote-sensing data gaps, Shapley additive explanation (SHAP) analysis to quantify factor contributions, and the Mann-Kendall test to detect trend breakpoints. Results show the total ITP lake area increased from 2.69 & times; 10(4) to 3.66 & times; 10(4) km(2) (36% growth) over 50 years: it remained relatively stable in 1972-1995, and entered a phase of marked expansion from 1996 (a breakpoint for annual mean area identified by Mann-Kendall test), and slowed after 2006 (a 2006 breakpoint for October area). Two large lakes-Selin Co and Ayakkum Lake-saw the largest area increases, exceeding 300 km(2) (+738 km(2) and + 520 km(2) respectively). A total of 433 new lakes formed over the period, with a combined area of 1255 km(2) (accounting for 12.4% of total ITP lake expansion), mostly concentrated in northern ITP. Notably, temporal scale dictates the dominant drivers of lake dynamics: subsurface runoff dominates at the monthly scale (with positive SHAP contributions when subsurface runoff > 10 mm), while energy factors (e.g., downward surface thermal radiation) are the primary drivers at the annual scale. This study clarifies the multi-scale mechanisms of ITP lake dynamics, providing robust data support for regional water resource management.

青藏高原（TP）素有“第三极”与“亚洲水塔”之称，其湖泊是气候变化的关键指示器。尽管近几十年来大多数青藏高原湖泊显著扩张，但高时间分辨率数据的缺口与尺度依赖的驱动机制仍不明晰。研究聚焦1972—2021年羌塘高原（高原内陆区，ITP）湖泊，采用基于自注意力机制的时间序列插补深度学习模型（SAITS）填补遥感数据缺口，利用SHAP（Shapley加性解释）分析量化因子贡献，并采用Mann-Kendall检验识别趋势突变点。结果表明：50年间羌塘内陆区湖泊总面积由2.69×10⁴ km²增至3.66×10⁴ km²（增长36%）——1972—1995年面积相对稳定，自1996年（Mann-Kendall检验识别的年均面积突变点）起进入显著扩张阶段，并在2006年（10月面积突变点）后增速放缓。色林错（Selin Co）与阿雅克库木湖（Ayakkum Lake）面积增量最大，分别达+738 km²与+520 km²（均超过300 km²）。期间共有433个新湖泊形成，合计面积1255 km²（占内陆区湖泊总扩张面积的12.4%），主要集中于羌塘北部。值得注意的是，时间尺度决定湖泊动态的主导驱动因子：月尺度上地下径流占主导（当地下径流 > 10 mm时SHAP贡献为正）；而年尺度上能量因子（如地表向下热辐射）是主要驱动。本研究厘清了羌塘内陆区湖泊动态的多尺度机制，为区域水资源管理提供了可靠的数据支撑。

（来源：Journal of Hydrology 2026  DOI: 10.1016/j.jhydrol.2026.135585）