Studying the spatial and temporal water distribution in the Lake Baikal Basin, which hosts the freshwater lake with the most water storage in the world, is essential to understand the water resources and environment of the basin and its impact and influence in terms of climate change and disaster prevention and mitigation. The basin spans two countries, Russia and Mongolia, which, along with its vastness, makes it challenging to accurately automate the acquisition of large-scale and long-term series data. The Google Earth Engine (GEE) is capable of processing large amounts of remote sensing imagery but does not support the computation and application of deep learning models. This study uses a combination of local deep learning training and GEE cloud-based big data intelligent computing to empower GEE with deep learning computing power, enabling it to rapidly automate the deployment of deep learning models. Visible light, near infrared (NIR), modified normalized difference water index (MNDWI), short-wave infrared 1 (SWIR1), linear enhancement band (LEB), and digital elevation model (DEM), which are more sensitive to water bodies, were selected as input features, along with the optimized input features of the existing pixel-based convolutional neural network (CNN) model. This method corrects the initial water labels from the Landsat quality assessment bands to reduce the time cost of manually drawing the labels and improving the classification accuracy of the water bodies. On average, only 1–2 h are required to generate the results for each water body product for each period in Lake Baikal Basin. The extraction of water bodies from the Lake Baikal Basin was achieved for nine yearly periods between 2013 and 2021. The validation accuracy was 92.9 %, 92.7 %, and 92.4 % for the three years 2013, 2017 and 2021, respectively. The results showed that the mean area of water bodies in the basin was 37,500 km² and that the area of water bodies in the basin fluctuated without significant change from 2013 to 2021. This study provides methodological support for the continuous monitoring and assessment of water body dynamics at more catchment scales and other large scenarios.

研究贝加尔湖流域的时空水分布，是世界上储水量最大的淡水湖，了解该流域的水资源和环境及其对气候变化和气候变化的影响和影响至关重要。防灾减灾。该盆地横跨俄罗斯和蒙古两个国家，加上其幅员辽阔，使得准确自动化获取大规模和长期系列数据具有挑战性。Google Earth Engine（GEE）能够处理大量的遥感影像，但不支持深度学习模型的计算和应用。本研究采用本地深度学习训练与GEE云端大数据智能计算相结合，为GEE赋能深度学习算力，使其能够快速自动部署深度学习模型。可见光、近红外（NIR）、修正归一化差分水体指数（MNDWI）、短波红外1（SWIR1）、线性增强带（LEB）、数字高程模型（DEM），对水体更敏感，与现有的基于像素的卷积神经网络 (CNN) 模型的优化输入特征一起被选为输入特征。该方法对来自Landsat质量评估波段的初始水体标签进行修正，以减少人工绘制标签的时间成本，提高水体分类精度。平均而言，生成贝加尔湖盆地每个时期每个水体产品的结果只需要 1-2 小时。贝加尔湖流域水体的提取在 2013 年至 2021 年的 9 个年度期间实现。2013 年、2017 年和 2021 年三年的验证准确度分别为 92.9%、92.7% 和 92.4%。结果表明，流域内水体平均面积为3.75万平方公​​里²、2013-2021年流域水体面积波动无明显变化。本研究为更多流域尺度等大场景水体动态的持续监测与评估提供了方法论支持。