This April (2024), a paper in Water Resources Research caught the attention of our Blog Committee. Intrigued by its innovative findings using a mesoscale hydrology model, we couldn’t resist going deeper. We had the pleasure of discussing this research with Pallav Shrestha, the lead author, and a prominent researcher at the Helmholtz Center for Environmental Research in Germany. Here’s an inside look at our fascinating conversation and the insights we uncovered.

Please introduce yourself to our readers.

I am Pallav Kumar Shrestha. I come from Nepal and currently reside in Germany. I joined the lab of Luis Samaniego at UFZ in 2017 as a PhD researcher, focusing on locally relevant flood forecasting in managed river basins at a global scale. I am one of the active developers of the mesoscale hydrological model (mHM) and the lead developer of the SCC river network upscaling technique and mHM’s reservoir module. My PhD journey is finally coming together with the acceptance of two recent publications: one in Nature Communications and another in Water Resources Research. Today, I’ll delve into the latter.

What led you to integrate a random forest model with the mesoscale Hydrological Model in your study?

While developing the reservoir module in mHM, we were able to satisfactorily represent all aspects of reservoirs based on physics, except for one: the water demand. Demand is a complex human response, highly discontinuous, and is therefore less suited to being modeled as a continuous function like other hydrological processes. For this reason, we hypothesized and demonstrated that machine learning techniques such as random forest could be available option to estimate demand. It is important to note that the random forest modeling exercise was external to mHM, and the fitted demand functions as model input.

Could you explain the significance of the Kling-Gupta Efficiency improvement noted in your study?

Sure. The incorporation of reservoirs in large-scale modeling applications is not a new concept. However, the improvement in Kling-Gupta Efficiency (KGE) for streamflow observed in our study, compared to the naturalized flow, is significantly higher than those reported in previous literature. We succeeded in matching the sub-seasonal details of the observed hydrograph, whereas the majority of previous studies focused only on matching seasonality. The improved KGE implies that modelers can now use mHM to represent reservoir regulation more accurately in their large-scale streamflow simulations.

What were the most challenging aspects of this research?

One of the challenges of this research was software development, including strategies for process representation, parameterization, debugging, testing, and, when things didn’t work, starting over from scratch. All this was very time-consuming and truly tested our resolve. During development, the most challenging aspect by far was getting the overall reservoir regulation right. Hours of meetings with Luis, Stephan, and Olda led to multiple strategies, some of which were even presented at international conferences. Looking back, most of these ideas never made it into this paper, which may seem inefficient, yet, as I learned, is an integral part of research. Yuval Harari puts it well: ‘The pursuit of truth is not feasible without the courage to waste time.’

How do you think your findings could influence future policies on water management and reservoir operations?

The random forest model for demand, which is based on time and meteorology, could prove valuable in hydrological forecasting. It is sensible to train the model recursively, limiting its application to sub- decadal horizons, as regulatory protocols may evolve over longer periods. Such models circumvent the intricacies of information sharing and should be leveraged, alongside satellite data, to foster regional cooperation for reservoir operations in transboundary basins. Representations that refine evaporation and scalable inflow hold promise for enhancing regional monitoring systems for reservoirs in drought-prone regions like the Iberian Peninsula. Exploring the possible relationship between reservoir disruptivity and ecological disturbance downstream allows for the quantification of impacts, enhancing the policy design of ‘eco-friendly reservoirs.’

In your paper, you mention the sensitivity of reservoir shape on streamflow and evaporation; could you elaborate on the practical implications of these findings?

Our experiment revealed that reservoir shape has minimal impact on streamflow but significant consequences on evaporation. Practically speaking, global models evaluated solely on streamflow may incur large errors in internal states (e.g., lake evaporation) if not properly checked. In such cases, optimizing model parameters could result in values that compensate for these errors—a classic case of ‘right results for wrong reasons.’ It is thus imperative to have a meaningful representation of reservoir shape and to check internal states in the model, such as evaporation, in addition to integration variables like streamflow.

You tested the Lake Module across 31 reservoirs. How did the geographical diversity of these locations impact your results?

Although our study did not include a detailed analysis of the impact of geographical diversity on the results, we did examine the role of climatic regions and physical characteristics of the reservoirs on evaporation. As anticipated, evaporation proved to be a significant component of the water balance in reservoirs located in arid climates and those with flatter shapes. Interestingly, the flattest and the narrowest reservoirs in our experiment were in humid and semi-arid climatic regions, respectively. The results indicated that evaporation from the flattest reservoir was the highest in the experiment, while evaporation from the narrowest reservoir was the lowest among all the reservoirs examined. Thus, among the factors studied, reservoir shape exerted a greater influence on evaporation than climate.

Your research identified that 30% of the non-consumptive hydropower reservoirs in the GRanD dataset are non-disruptive. How do you foresee this influencing global hydrological models?

It is estimated that there are 16.7 million reservoirs worldwide (Lehner et al., 2011), of which 38,000 are already georeferenced in databases like the GOODD (Mulligan et al., 2020), a number that is expected to increase overtime. Including every available reservoir in global streamflow simulations, given the parameterized regulation and iterative lake water balance schemes, would exacerbate an already computationally costly process. Our findings enable ‘disruptivity profiling’ of reservoirs, aiding large-scale modelers in prioritizing reservoirs for maximum streamflow regulation coverage within their computational budgets. This practice could help constrain the digital carbon footprint of computational hydrology, crucial in the era of hyperresolution modeling (Nature Computational Science Editorial Board, 2023).

How did the use of different shape approximations influence the accuracy of reservoir evaporation estimates in your simulations?

Reservoirs, impounded on natural terrain, depend on the instantaneous surface area for evaporation calculations. Without surveyed bathymetry, estimates of how surface area changes with depth are necessary. The common assumption of a rectangular prism, which considers the surface area constant at its maximum value, significantly overestimates evaporation. The ReGeom, a global bathymetry estimation dataset, also falls short as it only matches the maximum surface area and neglects the crucial depth-area profile essential for accurately estimating evaporation. Imagine a pyramid-shaped ice-cream cone, halved to reveal a triangular cross-section. This ‘half-pyramid’ approximation of natural terrain yielded the most accurate simulations across a variety of reservoirs due to its realistic depth-area profile.

In light of your research, what are the next steps in improving the representation of reservoirs in large-scale hydrological models?

First, we need innovations and a united effort to develop bathymetry datasets corresponding to global reservoir inventories, such as the GRanD. Although this paper proposes a globally applicable reservoir shape, relying solely on assumptions is outdated given our access to satellite technology. Second, as highlighted in this paper, the common methods for estimating lake evaporation are primitive at best. For example, the reflectivity (also known as albedo) of the water surface, which plays a crucial role in reservoir evaporation, varies spatially and temporally. Although easy to estimate, it is often assumed constant, which we have demonstrated is not the case. Finally, current large-scale models typically treat reservoirs within the same grid either partially (as ‘major’ and ‘minor’ reservoirs) or avoid this setup altogether. Our forthcoming publication on the subgrid catchment conservation (SCC) method promises to address this limitation.

Interviewed by Kyungdoe Han. This interview is part of the YHS Research ‘Hylights’ series, which introduces interesting and outstanding work by early-career scientists. The selection criteria are not set in stone, but reasons for selecting work may include the novelty and relevance of findings, enjoyment of reading, unique collaborations, media coverage, and generated controversy. Selected works will be provided with a brief written or video interview with the first author(s). Tips can be sent to young.hydrologicsociety@gmail.com.

References
Lehner, B., Liermann, C. R., Revenga, C., Vörösmarty, C., Fekete, B., Crouzet, P., … & Wisser, D. (2011). High‐resolution mapping of the world’s reservoirs and dams for sustainable river‐flow management. Frontiers in Ecology and the Environment, 9(9), 494-502.

Mulligan, M., van Soesbergen, A., & Sáenz, L. (2020). GOODD, a global dataset of more than 38,000 georeferenced dams. Scientific Data, 7(1), 31.

Nature Computational Science Editorial Board. (2023). The carbon footprint of computational research. Nat Comput Sci 3, 659 (2023). https://doi.org/10.1038/s43588-023-00506-2

Shrestha, P. K., Samaniego, L., Rakovec, O., Kumar, R., Mi, C., Rinke, K., & Thober, S. (2024). Toward improved simulations of disruptive reservoirs in global hydrological modeling. Water Resources Research, 60(4), e2023WR035433.