Accurate runoff prediction in data-poor catchments is important for water resource management, flood mitigation, environmental protection, and other tasks. One possible solution is to transfer a runoff prediction model constructed by using a machine learning model for gauged catchments to data-poor catchments. However, the transfer of runoff prediction model must consider the comprehensive spatiotemporal similarities between the catchments; otherwise, the transfer performance can be massively uncertain. Therefore, to improve the accuracy of runoff prediction and eliminate the uncertainty in identifying the differences between catchment environments, this paper proposes a novel measurement approach of comprehensive spatiotemporal similarity. This approach measures the similarities among catchments by fully considering which of the various catchments' spatiotemporal attributes can better represent the geographical similarity. Then, according to the similarities between the catchments, a runoff prediction model trained in gauged catchments is transformed for the most similar data-poor catchments to predict the runoff and the transfer performance is analyzed. To this end, a runoff prediction model is built using a gated recurrent unit (GRU) network based on the CAMELS catchments data set. A framework to extract the comprehensive spatiotemporal features of catchments is designed using three autoencoders. The catchments' similarities can be measured, further, and their spatiotemporal attributes determined once a measurement model of comprehensive spatiotemporal similarity is constructed. Finally, the transfer performance of the GRU runoff prediction model based on comprehensive spatiotemporal and other geographical similarities is evaluated and analyzed. The experimental results demonstrate that the proposed method outperforms comparable approaches.
Keywords: Comprehensive spatiotemporal similarity; GRU runoff Model; Model transfer; Runoff prediction.
Copyright © 2022 Elsevier Ltd. All rights reserved.