科研之友
微信
新浪微博
Facebook
分享链接
摘要
This study develops a novel approach to data-driven hydrological modeling. The approach adopts the feature representation technique in computer vision to effectively exploit spatial information contained in time-variant input data fields and seamlessly fuse multisource information via machine learning. The new approach overcomes a major limitation of existing approaches in which the spatial heterogeneity of input variables cannot be sufficiently accounted for. The approach is applied to predict the streamflow in a watershed on the northern margin of the Qinghai-Tibetan Plateau, and its performance is compared with various data-driven and process-based models. The major findings are as follows. First, the new approach represents a general framework for the fusion of multisource spatiotemporal data for hydrological modeling and demonstrates great potential to incorporate fast-growing environmental big data. Second, the new approach demonstrates satisfactory short-term forecasting, long-term simulation, and transfer learning performances and is promising for addressing predictions in ungauged basins. Third, the predictors, including precipitation, temperature, leaf area index, and historical streamflow, play markedly distinct roles in modeling streamflow with the novel approach. Finally, topographic information is not a necessary model input in the proposed approach because spatial patterns can be well embodied by other inputs (e.g., temperature) that have high similarities with topography. This study represents the first attempt to bring computer vision into data-driven hydrological modeling and may inspire future studies in this promising direction.
- 出版日期2018-12
