Advancing Synergistic Knowledge Generation through Machine Learning and Hydrology Science

Abstract

Our society has experienced significant technological advancements over the last few decades, particularly in science. In hydrology, the two main modeling approaches are facing similar situations. The process-based approach has achieved larger scales and better resolution, while the data-based approach has demonstrated unprecedented performance. This creates a significant opportunity for synergistic learning between both approaches that could enhance the field of hydrology, but currently, these opportunities are not fully explored. For example, the new knowledge extracted from machine learning models is often not interpreted in a hydrological context, focusing solely on performance. Another example occurs each time machine learning models start training from scratch, disregarding decades of acquired hydrological knowledge. Thus, there is a remarkable opportunity for synergistic learning, but we must first push to integrate both approaches. This dissertation explores steps toward synergetic knowledge generation using machine learning models alongside the transferability of hydrological knowledge encoded in process-based models. The first contribution in this direction is developing a hydrologically interpretable machine learning architecture inspired by a water reservoir (HydroLSTM). From this architecture, we learned how machine learning can encode the hydrological dynamics of a catchment using specific pattern weights. The second contribution shows that these patterns can be regionalized based on catchment attributes, resulting in consistent spatial distributions of dynamically similar catchments and distinctive hydrological behaviors in the state variable and output gate. In the third experiment, we explore the use of hydrological knowledge to augment machine learning models under sparse and biased data. We pre-trained a machine learning model using results from a process-based model to capture hydrological relationships. This pre-trained model is then trained on observational data. The resulting model incorporated more information about snow water equivalent (SWE) and yielded better spatial generalization than a model trained solely on observations. These findings highlight the significance of hydrological knowledge, particularly when data is insufficient for directly training a machine learning model. Interesting connections and mutual benefits between hydrology and machine learning were revealed in this dissertation. For instance, the potential to extract hydrological knowledge from a machine learning model can advance hydrological discoveries. Moreover, the transferability of knowledge from hydrological models presents new opportunities for improving model performance. These insights suggest that the synergy between our knowledge and machine learning can lead us to a new era in hydrology science.