Monitoring of lake water level (WL) changes with satellite altimeters is pivotal in assessing the dynamics of hydrological ecosystems. However, the spatial coverage, temporal interval, and quality degradation of altimeter data limit the continuity of the measurements. In thisarticle, a learning-based framework is proposed for the reconstruction of WLs for inland lakes and reservoirs. This is achieved by learning the relationship between lake WLs and the related hydrological and climate variables employing the deep learning models. By introducing hydrological knowledge into the data-driven learning framework, the lakes are first clustered into several groups for training and prediction considering the spatial homogeneity and heterogeneity of water cycling process among multiple lakes. Second, for each cluster category, the number of WL training samples is augmented using the empirical function fitted with lake level-area pairs, and the augmented samples are used in the pretraining process to improve the accuracy of the deep learning model simulation. The obtained models are used for estimating the missing WLs and to construct a consecutive 192-month WL dataset (2003-2018) for the 14 lakes (>20 km2) in the Altai Mountains. The typical multiple layer perceptron and deep belief network models are tested. Validation indicates that the proposed method performs well in WL reconstruction in the case of a large proportion of missing data. Moreover, the performance of learning-based models can be effectively improved by introducing the idea of spatial clustering and pretraining. The comparative tests also show that the proposed method outperforms the traditional level-area fitting methods. © 2008-2012 IEEE.
