Bathymetric mapping for an accurate estimation of stored water volume in drying lakes is a key information for an effective monitoring of their recession or restoration status. Extraction of bathymetry in shallow saline lakes using remote sensing techniques has always been challenging due to the complex influences imposed by the physical properties of substrate and the spatial variability of salinity. In this study, we developed a machine learning-based model to quantify the implicit, non-linear relationship between water depth and surface reflectance by leveraging extensive in-situ data and high-resolution satellite imagery. We trained and tested the learning model in the hyper-saline Lake Urmia (LU), which faced catastrophic drying over the past two decades. To this end, we used Landsat-8 imagery and 32,984 hydrography data points surveyed by the Urmia Lake Restoration Program (ULRP) from 2017 to 2020 during six stages. To enhance the model accuracy, we tuned the model inputs by optimizing the spectral information and clustering in-situ data from stages with similar meteorological conditions into three classes. The results demonstrated the high accuracy of the developed intelligent model as evidenced by R² = 0.8 ∼ 0.9 and RMSE = 7.8 ∼ 17.9 cm for the three models. We found that the average water depth in the LU was increased from 0.43 m in September 2018 to 2.00 m in May 2020. In particular, the lake water volume in May 2020 was 3.6 times greater than that in February 2019, which marks a remarkable shift in the LU restoration. Dynamic bathymetric maps also witnessed considerable salt dissolution taking place across the lake during this period. Finally, we extracted the LU level-area relationship by processing 172 Landsat images between 1984 and 2020, which was validated against the field data surveyed along the lake water boundary in 2019. The results indicated that the level-area relationship follows a dual linear relationship separated at the water level of 1271.31 m.

用于准确估计干涸湖泊中储存水量的测深图是有效监测其衰退或恢复状态的关键信息。由于底物物理特性和盐度空间变异性的复杂影响，使用遥感技术提取浅盐湖水深一直具有挑战性。在这项研究中，我们开发了一个基于机器学习的模型，通过利用广泛的现场数据和高分辨率卫星图像来量化水深和表面反射率之间的隐式非线性关系。我们在高盐度的乌尔米亚湖 (LU) 中训练和测试了学习模型，该湖在过去的二十年中面临着灾难性的干旱。为此，我们使用了 Landsat-8 图像和 32, 乌尔米亚湖恢复计划（ULRP）从 2017 年到 2020 年在六个阶段调查了 984 个水文数据点。为了提高模型的准确性，我们通过优化光谱信息并将来自具有相似气象条件的阶段的现场数据分为三类来调整模型输入。结果证明了开发的智能模型的高精度，如 R 所证明的 三个模型的² = 0.8 ∼ 0.9 和 RMSE = 7.8 ∼ 17.9 cm。我们发现，LU 平均水深从 2018 年 9 月的 0.43 m 增加到 2020 年 5 月的 2.00 m。特别是 2020 年 5 月的湖水量是 2019 年 2 月的 3.6 倍，这标志着一个显着的转变在LU恢复中。动态测深图还见证了在此期间整个湖泊发生的大量盐分溶解。最后，我们通过处理 172 幅 1984 年至 2020 年的 Landsat 影像提取 LU 水平-面积关系，并与 2019 年沿湖水边界调查的野外数据进行验证。结果表明水平-面积关系遵循双重线性关系在水位1271.31 m处分离。