快速准确地获取地表水分布情况有助于开展水资源管理工作和预防洪涝灾害的发生；掌握长时期地表水时空分布规律对研究气候变化和指导人类活动具有重要意义；地表水变化不仅受到气候因素影响，还可能受到地形植被等多种潜在因素的影响，总结地表水驱动因素的时空影响差异性有助于全面了解驱动因素作用机制，预测未来地表水变化情景。青海省三江源区是亚洲重要的生态安全屏障，但同时也是气候变化敏感区，地表水变化错综复杂和难以调控。如何方便准确地获取三江源区地表水范围，全面分析地表水历史变化规律，多尺度揭示地表水驱动因素影响机制，是实现水资源可持续发展的关键。

本文使用Landsat系列遥感影像，基于谷歌地球引擎（Google Earth Engine, GEE）平台，探究不同特征信息与深度学习模型结构对地表水提取精度的影响，构建了在线自动化准确提取地表水范围的方法，并获取了1990-2020年三江源地表水分布范围；从时空尺度综合分析季节性、年度以及多年性地表水变化规律，探讨不同类型湖泊的时空格局演变过程；揭示了气象、植被、地形和土壤理化性质等驱动因子对年度地表水影响的空间异质性。主要结论如下：

（1）在线自动化准确提取地表水方法

针对构建的10种具有不同输入特征或网络结构的深度学习模型展开离线本地训练与精度评价，结果表明使用红绿蓝波段和AWEI_sh指数作为输入的轻量化U-net++模型表现最佳，总体Kappa系数达到98.87%；将该最佳模型按相同参数和结构迁移至GEE云端实现了免费的在线自动提取水体过程，经测试单次提取整个三江源范围水体的时间不超过3分钟。

基于百分比合成法，确定了以5%作为百分位数合成的影像可以提取三江源地表最大水体范围，经参考数据拟合验证，确定了以65%作为百分位数合成的影像可以提取三江源地表年度水体范围，两者通过逻辑运算即可获得三江源季节性水体；基于频率法，确定了1990-2020年来至少在29年里始终被观测为年度水体的地表水即为三江源多年性水体。

（2）三江源地表水与地表湖泊时空变化规律

从1990-2020年时间序列尺度上，三江源季节性和年度水体面积呈现上升趋势，31年来分别增加了4507.90 km²和1354.61 km²，年度水体和季节性水体在不同时段呈现不同的变化趋势。从空间尺度上，1990与2020年的地表水状态转换过程剧烈，新增年度水体的面积为2375.56 km²，新增季节性水体面积为6344.17 km²，消失的年度水体面积为1021.07 km²，消失的季节性水体面积为1836.64 km²，年度水体转季节性水体的面积为448.09 km²，季节性水体转年度水体的面积为469.00 km²，持续存在的年度水体面积为8635.29 km²，持续存在的季节性水体面积为538.82 km²；此外三江源多年性水体总面积为7251.83 km²，并集中分布于三江源区西北部湖泊群、中部以鄂陵湖、扎陵湖为主的湖泊群、东北部的龙羊峡水库和部分青海湖区以及大型主干河流（如长江上游的通天河）等位置。

根据湖泊提取分类及景观格局尺度分析结果，微小湖泊广泛分布于西北部、中部湖泊群，小型湖泊基本分布于三大湖泊群；冰川补给湖和中型湖泊的面积呈下降趋势，热融湖塘呈现出较明显地上升趋势；冰川补给湖的景观格局存在空间差异，西北部冰川补给湖的数量明显下降，西南侧冰川融水对湖泊的补给存在不均衡的现象。

（3）三江源区年度地表水长时序驱动因素分析

从1990-2020年驱动因素平均结果来看，降水、气温和NDVI是局部驱动因子，对地表水影响的空间异质性最强，海拔则为全局驱动因子；降水、气温、风速的对地表水体具有正向影响，相对湿度、NDVI、海拔、坡度、土壤有机质含量均呈现负向影响。从三个突变年份（1998、2001、2006年）的驱动因素结果来看，降水始终呈现正向影响，但总体强度在3个年份中逐渐减小；相对湿度对地表年度水体的显著影响几乎是负向的，但对三江源中部的湖泊群附近为正向影响；气温始终呈现正向影响，对中部湖泊群的正向影响最为明显；NDVI对地表水的影响几乎是负向的，并且对中部湖泊群的影响最为明显；坡度在地表水聚集的流域附近呈现负向影响，并且负向影响强度逐渐减弱。

The accurate and rapid acquisition of surface water distribution is crucial for water resource management and flood disaster prevention. Understanding the long-term spatiotemporal distribution patterns of surface water is significant for studying climate change and guiding human activities. Surface water changes are influenced not only by climatic factors but also by potential factors such as topography and vegetation. Summarizing the spatiotemporal differences in the drivers of surface water helps in comprehensively understanding the mechanisms of these driving forces and predicting future changes in surface water scenarios. The Three-river Source in Qinghai Province is an important ecological security barrier in Asia, yet it is also a climate-sensitive region where surface water changes are complex and challenging to manage. Developing a convenient and accurate method to obtain the range of surface water in Three-river Source, thoroughly analyzing the historical patterns of surface water, and revealing the mechanisms of multiscale drivers are key to achieving sustainable water resource development.

This study utilizes Landsat series remote sensing images and the Google Earth Engine (GEE) platform to explore the impact of different feature information and deep learning model structures on the accuracy of surface water extraction. An automated and accurate method for extracting surface water online was developed, and the distribution of surface water from 1990 to 2020 in Three-river Source was obtained. The study comprehensively analyzed the seasonal, annual, and permanent changes in surface water from spatiotemporal scales, discussed the evolutionary process of different types of lakes, and revealed the spatial heterogeneity of meteorological, vegetation, topographical, and geological drivers on annual surface water. The main conclusions are as follows:

(1) Automated and accurate online method for extracting surface water

Ten deep learning models with different input features or network structures were developed for offline local training and accuracy assessment. The results show that the lightweight U-net++ model using RGB bands and AWEIsh index as inputs performed the best, with an overall Kappa coefficient reaching 98.87%. This optimal model was transferred to the GEE cloud for free online automated water body extraction, and the entire Three-river Source could be processed in less than three minutes per extraction.

Using the percentile composite method, the image composed at the 5th percentile was determined to extract the maximum surface water body range of the Three-river Source, and the image composed at the 65th percentile was verified to extract the annual water body range of Three-river Source. Both were obtained through logical operations to achieve the seasonal water body. Based on the frequency method, the surface water observed as annual water bodies in at least 29 of the years from 1990 to 2020 was identified as the permanent water bodies of Three-river Source.

(2) Spatiotemporal patterns of surface water and lakes in Three-river Source

From the time series scale of 1990-2020, the areas of seasonal and annual water bodies showed an increasing trend, expanding by 4507.90 km² and 1354.61 km², respectively. The trends varied across different periods. From a spatial perspective, the transformations of surface water status between 1990 and 2020 were dramatic. Newly added annual water bodies covered an area of 2375.56 km², and seasonal water bodies increased by 6344.17 km². The disappeared annual and seasonal water bodies covered 1021.07 km² and 1836.64 km², respectively. The area transitioning from annual to seasonal water bodies was 448.09 km², and from seasonal to annual was 469.00 km². The persistently existing annual and seasonal water bodies covered 8635.29 km² and 538.82 km², respectively. Additionally, the total area of permanent water bodies in Three-river Source was 7251.83 km², primarily located in the northwestern lake clusters, central areas around Eling Lake and Zhaling Lake, the northeastern Longyangxia Reservoir, parts of Qinghai Lake, and major tributaries such as the upper reaches of the Yangtze River.

Based on the classification and landscape pattern scale analysis of lakes, tiny lakes were widely distributed in the northwestern and central lake clusters, and small lakes were mainly found in the three major lake clusters. The area of glacier-fed lakes and medium-sized lakes showed a decreasing trend, while thermokarst ponds exhibited a notable increase. There were spatial differences in the landscape patterns of glacier-fed lakes; the number in the northwest significantly decreased, and the glacier meltwater supply to lakes on the southwest side was uneven.

(3) Long-term analysis of annual surface water drivers in Three-river Source

From the average results of drivers from 1990 to 2020, precipitation, temperature, and NDVI were local driving factors with the strongest spatial heterogeneity, while elevation acted as a global driver. Precipitation, temperature, and wind speed had a positive influence on surface water bodies, whereas relative humidity, NDVI, elevation, and slope showed negative effects. In three anomaly years (1998, 2001, and 2006), precipitation consistently had a positive influence, but its overall intensity gradually decreased across the years. Relative humidity almost invariably had a negative impact on annual water bodies, but it positively affected the lake clusters in the central of Three-river Source. Temperature consistently had a positive impact, with the most significant positive effects observed in the central lake clusters. NDVI's impact on surface water was almost entirely negative, and its influence was most notable in the central lake clusters. Slope showed a negative impact near watersheds where surface water accumulated, and the intensity of this negative impact gradually weakened.