Performance report: Le Gardon – Pont de Russan catchment

In our latest Performance Report, we assess the effectiveness of BWI’s discharge forecast model for the Gardon River, specifically at the Pont de Russan outlet. Covering an area of 1,533.7 km², this catchment, located northwest of Nîmes, in southern France, features 362 reaches and 4 dams. The region typically experiences dry conditions with low baseflow (2-5 m³/s) but is vulnerable to sudden heavy rainfall and flash flooding. The model demonstrates strong overall performance (NNSE 0.81), accurately simulating baseflow, flood detection, and flood recession. This performance is crucial for managing water resources, supporting early warning systems, and aiding the design of critical flood prevention infrastructure.

With the latest UN water data revealing that the world’s rivers are drying at their fastest rate in three decades, the impacts of climate change are increasingly undeniable. Across vast regions of North, Central, and South America, as well as Asia and Oceania, well-known rivers like the Amazon and Mississippi are hitting record lows. Meanwhile, other areas face a starkly different reality, with rising water levels from more floods, storms, and intense downpours (BBC, 2024).

This makes water a crucial indicator of climate change, with weather extremes accelerating the hydrological cycle (BBC, 2024). The catchment area of Le Gardon river, France, explored in our latest performance report, is normally dry but notorious for flash floods after heavy rains. The sections that follow examine the performance of BWI’s semi-distributed hydrological model, which integrates machine learning techniques, to monitor and forecast river discharge. Specifically, we focus on Le Gardon at Pont de Russan, located just north of Nîmes in France’s Gard department.

Catchment Overview

The Gardon river catchment is situated in the Occitanie region in southern France, flowing primarily through the Gard department northwest of Nîmes. It is a right-hand tributary of the Rhône, into which it flows after a journey of 127.6 km. The river is, therefore, a part of a complex hydrological network, having several sub-catchments feeding into the main river.

Its source is in the Cévennes mountains, a low mountain range with a 1699-meter peak, the Pic de Finiels, specific for rugged terrain and dense forests. Covering approximately 2,000 square kilometers, it contains two main upstream reaches, the Gardon d’Alès and the Gardon d’Anduze, and a single downstream reach.

The catchment area is defined by varied landscapes, ranging from mountainous areas to the plains, which significantly influence the water flows and erosion processes within the system. In addition, its Mediterranean climate, with hot, dry summers and wet, stormy autumns, further intensifies the variability of the erratic flow rates.

 

Validation Metrics

In hydrological modeling, the components of performance metrics like the Nash-Sutcliffe Efficiency (NSE) and the Kling-Gupta Efficiency (KGE) provide crucial insights into various aspects of discharge prediction. At BWI, we consider both NSE and KGE for model performance evaluation and we deploy the Normalized NSE (NNSE) to communicate the model performance evaluation with the stakeholders.

NSE primarily assesses the overall agreement between observed and simulated discharge. It does not have specific components like KGE, but it indirectly provides information on the peak flow accuracy, overall variability, and bias (indirectly) of the discharges produced by the hydrological model at the catchment outlet. It ranges from −∞ to 1 and evaluates how well the simulated data matches the observed data. NNSE adjusts the NSE value to make it more comparable across different catchments or periods by standardizing to provide indices expressed between 0 and 1.

NSE is particularly sensitive to peak flows. It emphasizes how well the model captures high discharge events, such as flash floods, which are critical in the La Gardon river basin. A high NSE value indicates that the model is accurately predicting peak discharges, reflecting the catchment’s response to extreme rainfall.

By comparing the squared differences between observed and simulated discharges to the variability in observed discharges, NSE indicates how well the model captures the overall variability of flows in the catchment.

Although NSE does not explicitly separate bias, a low NSE score could indicate that the model is consistently underestimating or overestimating the flow. It suggests that the simulated and observed discharges differ significantly, often due to bias in the model.

For a hydrological model, the values 0 < NNSE < 0.5 indicate a poor model performance, at NNSE > 0.67 the baseflow corelates well with the observation and we use it as a minimum threshold to deploy the hydrological model to forecast 10 days discharge predictions. At NNSE > 0.85 the peak flow shows good agreement with timing and magnitude of extreme event floods. NNSE of 1 represents a perfect model.

The performance of “Le Gardon – Pont de Russan” model is shown in the figure below. The catchment map of the outlet is shown on the left and the graphical comparison between observed discharges (in gold) and simulated discharges (in green) is shown on the right.

The catchment is in the north-west of Nîmes, France. The region is normally dry with a low baseflow (2-5 m3/s) at the outlet but is prone to sudden heavy rains and flash floods. In the discharge plot, in the training period (left of the red dotted line) shows sudden flood events that elevate the discharge from baseflow of 2-5 m3/s to 200-400 m3/s. In the test period, however, reached ~1400 m3/s at an instance. The flood events are well identified, with good recession matches by the model. The amplitude up to 400 m3/s are well matched whereas due to the rarity of floods above that magnitude, the peak magnitude is quite difficult to capture. Also, the weather data lacks the required resolution to accurately capture such peaks.

The model shows a high overall model performance (NNSE 0.81) at the selected outlet. The well-simulated baseflow, flood detection, and flood recession are useful for water resource management, early warning systems, and planning of critical flood prevention infrastructures. As the model gets more data on extreme events, it will also improve the capability of predicting such events.

Reference list:

Académie Pont du Gard (2014) “Le Gardon, l’enfant terrible”. Available at: https://www.academie-pontdugard.com/le-gardon-lenfant-terrible/ (Accessed 8 October 2024).

BBC Global News Podcast (2024) “Netanyahu vows to continue fighting as Israel remembers 7th october attacks”, [Podcast]. BBC. 08 October 2024. Available at: https://www.bbc.co.uk/sounds/play/p0jw7t9q (Accessed: 08 October 2024). Timestamp: 00:24:00

Hassan, D., Isaac, G.A., Taylor, P.A.,  Michelson, D. (2022) “Optimizing Radar-Based Rainfall Estimation Using Machine Learning Models” Remote Sensing, vol. 14, no. 20. Available at: https://www.researchgate.net/publication/364346865_Optimizing_Radar-Based_Rainfall_Estimation_Using_Machine_Learning_Models (Accessed 11 October 2024).