Climate change means water change. To adapt and plan, the world needs information on which to base water decisions, and yet most of the world’s rivers have no historical data records. Fewer still have readily available information about their current state and forecasted future. If we are to empower decision makers - conservation experts, water managers, energy operators, emergency responders, and every other person that touches water - in a future environment of water uncertainty, reliable and accessible data is crucial.
At Upstream Tech, we built HydroForecast to address this pressing need for water intelligence. We endeavor to create a global understanding of the past, present and future of water at a scale and level of accuracy never done before.
These river basins fall within the Kavango Zambezi Transfrontier Conservation Area (KAZA). KAZA was established in 2011 as the world’s largest transboundary protected region to help conserve wild expanses, stretching 106 million acres across Angola, Botswana, Namibia, Zambia, and Zimbabwe. Leaders of the five partner countries have a shared vision to protect wildlife, improve the wellbeing of local communities, and promote tourism. WWF is a conservation partner to KAZA and collaborates with the group to prevent poaching, conduct scientific research on wildlife, promote habitat protection, and identify opportunities for communities to manage and benefit from wildlife on their land. The Kwando and Upper Zambezi river basins are critical habitats within KAZA, supporting wildlife that depends on the ebbs and pulses of the rivers.
Little is understood about these basins' hydrology, making it difficult to plan and protect water for nature’s needs as human consumption grows and weather patterns shift. The WWF is working on the ground to bring together the stakeholders and experts needed to develop strategies and policies for protecting and managing the critical KAZA corridor.
We are using HydroForecast, Upstream Tech’s theory-guided machine learning model, to provide historical ungauged streamflow records at key locations in the Kwando and Upper Zambezi basins and produce ongoing seasonal forecasts at these same locations. HydroForecast offers a unique opportunity for WWF to use cutting edge artificial intelligence tools that can deliver hydrologic information, even in places without on-the-ground sensors.
Current and historical forecast results are available on our publicly accessible, interactive website and hosted in Microsoft's Planetary Computer Catalog. The Planetary Computer supports sustainable decision making by hosting a catalog of global environmental data with intuitive APIs and applications for public use. Through these channels, anyone can view historical and future streamflow in the Kwando and Upper Zambezi basins. We view this project as the first in a series in which HydroForecast sheds light on otherwise opaque hydrological conditions and outlooks in regions around the world.
Designing a New Process
Through a collaborative process, six locations in the Upper Zambezi and Kwando basin were identified as key locations for producing reanalysis streamflow records from HydroForecast. The main factors influencing these locations were the availability of in situ data for accuracy validation, the co-location with another modeling effort by Duke and Rhodes Universities, and the expert opinions from the WWF-Zambia team.
The process for producing reanalysis streamflow records at each location is summarized here:
1. Start with the HydroForecast core model trained on ~450 basins in North America to teach the model general hydrologic principles; this model utilized global input data sources.
2. Upper Zambezi: Tune a model in the two subbasins and validate using in situ observations.
- Trained on data from 1984 - 2015
- Validation period: 2015 - 2021
3. Compute standard goodness of fit metrics for assessing model performance.
- Kling-Gupta Efficiency (KGE)
- Nash Sutcliffe Efficiency (NSE)
- Bias (as a percent)
- Mean Absolute Error (MAE) in cubic meters per second (cms)
- Pearson’s correlation coefficient
What We Found
Starting in the Upper Zambezi. Using our core model trained across hundreds of basins in the U.S., we tuned a model to the Upper Zambezi catchments, where observed streamflow records were available. Figure 1 depicts the validation period from 2015-2021 in the Upper Zambezi BAR7 catchment. During this time, the model has never seen observations nor has it ever seen this time period in any other basin. Notice the alignment between the model’s mean prediction (dark purple) and the observations (blue) during the important peak events. The model captures the large flow events within the 90% confidence interval (shaded bands), and nails the timing of the peaks, nearly matching the magnitude in the higher flow years.
Creating Kwando basin predictions. We took this model tuned on the Upper Zambezi and applied it to the Kwando basin, and the result is a continuous daily time-series of our best estimates of streamflow from 1982-2022. Figure 2 illustrates the model compared with the gauge observations at the Kwando basin’s outlet at Kongola station.
Virtual gauge predictions. Gauge observations only exist at the Kongola station, so we used our model to create ‘virtual gauges’ of simulated streamflow over the historical period. In a place with limited and sparse data, these records could be used as a guide for understanding general flow patterns and ranges. In Figure 3, we show the predictions for the upper most headwater catchment, Sub Basin 1.
Applying the Findings
Anyone can view and download our historical reanalysis and seasonal forecasts, which are available on our public website and hosted in Microsoft's Planetary Computer Catalog. The data will be used by WWF, community decision makers, and other organizations on an ongoing basis. The historical virtual gauge streamflow records will be the basis for informing environmental flow policies that are currently under development. Ongoing seasonal forecasts will help organizations, local governments, farmers and all water users plan for the upcoming season.
Want to build something using this data? We want to hear about it! If you’re interested in learning more about HydroForecast’s capabilities and wish to view other forecast examples, visit our website or get in touch with us today.