Role of Machine Learning & Modelling in Fluvial Geomorphology

Tuesday June 27 11:20am-12:45pm

Leads: Adeyemi Olusola (York University) and Corey Dawson (Dalhousie University)

River dynamics are complex and difficult to predict. Machine learning, and new means of data collection (e.g., remote sensing, GIS) offer up new possibilities for understanding these processes, and ultimately improve channel restoration efforts. This session aims to highlight recent work that uses these approaches to understanding river dynamics (e.g., machine learning, combinations of machine learning and field-based observations, GIS solutions). We welcome those interested in furthering this conversation and looking to learn more about machine learning algorithms in fluvial geomorphology.

Keywords: Artificial intelligence, Modelling, Classification, Process-based Approaches, GIS and Remote Sensing Applications

Following the conference, presentations that have been made available will be linked here.

Click on the titles below to read the abstract for a given presentation. Titles are listed in presentation order.
Talk 1 - Predicting River Morphology Classification

Cody Kupferschmidt and Andrew Binns 

School of Engineering, University of Guelph, Guelph, Canada

Numerous classification schemes have been developed for categorizing rivers based on their morphologies and behaviour. One of the most well-known form-based morphological classification system was developed by Rosgen (1994, 1996) and uses dimensional properties such as entrenchment ratio, width to depth ratio, and sinuosity to group streams into eight main classes, with streams further classified into sub-classes using additional features such as bed material and channel slope. The Rosgen (1994, 1996) system is currently one of the most widely accepted classification systems used in North America, and has been adopted by several government agencies including the US Natural Resources Conservation Service and US Forest Service.

In the last decade, machine learning models known as deep convolutional neural networks (CNNs) have seen extensive use for performing image classification tasks (e.g. VGG-16, ResNet). While there is great potential and considerable interest in the use of machine learning techniques in the water resources field, implementations of deep learning techniques have been limited. Some research has suggested that a lack of benchmark datasets may be a limiting factor in the application of data science in geosciences.

The authors are not aware of any benchmarking datasets that currently exist for training machine learning models to perform river morphology classification tasks. To-date, the largest collection of labelled river morphology images appears to exist in the book Applied River Morphology (Rosgen, 1996), containing more than 250 images of rivers, each labelled according to their Rosgen class. In the present study, we curate a dataset, starting with the image collection from Rosgen (1996) and adding additional images that have also been labelled with their Rosgen class. The new dataset is used to train and test a CNN model for the task of predicting the Rosgen class of river images.

We present key findings from the study, and discuss challenges surrounding the classification of natural landscapes. Finally, we encourage Canadian practitioners to advocate for and participate in the creation of Canada-specific geoscience benchmarking datasets.

Rosgen, D.L. (1994). A Classification of Natural Rivers. Catena. 22(3): 169-199. doi:

Rosgen, D.L. (1996). Applied River Morphology. Pagosa Springs, Colorado: Wildland Hydrology.

Talk 2 - Predicting river discharge characteristics using deep learning algorithms

Adeyemi Olusola1, Samuel Ogunjo2 and Christiana Olusegun3

1Faculty of Environmental and Urban Change, York University, Toronto, ­­Canada

2Department of Physics, Federal University of Technology, Akure, Ondo State, Nigeria

3Faculty of Physics, University of Warsaw, Poland

Across West Africa, the River Niger is a significant source of freshwater for various uses, including irrigation, aquaculture, transportation, and hydropower. The river network plays a critical role in the socioeconomic development and stability within the region. Therefore, understanding the inherent dynamics and associated environmental flows is essential to maintain a healthy river ecosystem amidst varying hydro-political priorities. However, with the persistence of 5 drought conditions, threats of climate change, and associated teleconnections, there is a need to predict the flow of the River Niger system within short time frames. Hence, this study aims to predict the River Niger discharge at eight stations – Koulikoro, Dioila, Kirango, Douna, Mopti, Dire, Ansongo, and Niamey from 1950 – 1990 using deep learning algorithms. The analysis for this study was performed using the darts python library, while the learning of the river discharge was based on the past histories of the river discharge at each of the study locations. For the monthly records, the deep learning algorithm 10 model based on temporal convolutional networks (TCN) and long short-term memory (LSTM) performed better than recurrent neural networks (RNN) and gated recurrent units (GRU). We found that no one-size-fits-all model can be used to predict different annual discharge characteristics across the stations within the study area. However, TCN and LSTM performed superior in most characteristics based on root mean square error values. Low root mean square errors in the range 0 − 4 were observed in Diola stations for minimum annual river discharge.

Talk 3 - Performance Assessment of Permeable and Top-blocked Permeable Groins-in-series in Riverbank Protection and River Training, Using 2D/3D Hydrodynamic and Sediment Transport Modeling (TUFLOW): A Case Study along the Jamuna River, Bangladesh

Dr. Bahar SM1, Sadia Afrin Khan2 and Labiba Fairoze Prottyasha2

1AHYDTECH Geomorphic Ltd., Guelph, Ontario, Canada

2AHYDTECH Water resources (BD) Ltd., Dhaka, Bangladesh

Jamuna River is one of the largest braided rivers in the world. Failure of riverbank protection structures in the Jamuna River is a common phenomenon, mainly because interventions like concrete block/geo-bag revetments are not strong enough against morphological dynamics and shifting patterns of this unpredictable braided river system. While the cross-channel interventions like impermeable groins/spur dikes cannot sustain the associated extreme local scour beneath the structures. Therefore, this study develops 2D/3D hydrodynamic and sediment transport models using TUFLOW for a highly erosion- prone bank (average erosion rate being up to 145 meters/year) along the Jamuna River (Kalihati upazila, Bangladesh) and explores the performance of other river training structures like permeable groins (PG) and top-blocked permeable groins (TBPG) in- series, by analyzing the change in bed level, localized erosion/sedimentation, velocity, and shear stress at multiple monitoring locations, for several alternative scenarios and temporal durations. The 2D/3D hydrodynamic and sediment transport models were calibrated with respect to the observed water level data at Bahadurabad gauge station and annual sediment load of the Jamuna River, respectively. This study has done extensive sensitivity analysis of sediment transport parameters, including suspended load, bedload, erosion, deposition and bank slumping. The results show that compared to the initial condition, the riverbank can erode over 90-meters horizontally and over 28 meters vertically at several locations after one year if no bank protection structures are installed, but the PGs and TBPGs in-series entirely protect the riverbank, and even results in sedimentation along the bank. Additionally, during the monsoon period (July-August), the maximum and average flow velocity with groins reduce spatially compared to the without groins scenario (e.g., the reduction is 7.78% and 18.54%, respectively, at the downstream of the first groin). These observations indicate that adapting the PG or TBPG implementation in Jamuna has great potential for riverbank protection, land reclamation, and local navigation improvement.

Talk 4 - Integrated Watershed Modelling for Evaluating Cost Effectiveness of Natural Infrastructure for Mitigating Road Washouts in Parkland County, Alberta

Wanhong Yang1, Shawn Shao2, Yongbo Liu3, Tien Weber4, Rhonda King5, Mary Ellen Shain6 and Krista Quesnel7

1Department of Geography, Environment and Geomatics, University of Guelph, Guelph, Ontario, Canada

2Esri Canada Limited, Ottawa, Ontario, Canada

3Environment and Climate Change Canada, Burlington, Ontario

4Department of Renewable Resource, University of Alberta, Edmonton, Alberta, Canada

5ALUS Canada, Kiscoty, Alberta, Canada

6North Sasketchewan Watershed Alliance, Edmonton, Alberta, Canada

7Parkland County, Stony Plain, Alberta, Canada

Flooding related road washouts is an ongoing infrastructure concern in Parkland County, Alberta. The traditional engineering approach is to augment drainage capacity such as culvert size to address this issue. In this study, Parkland County collaborated with ALUS Canada, North Saskatchewan Watershed Alliance and the University of Guelph to explore the possibility of implementing natural infrastructure including riparian and wetland restoration to reduce flow peak for mitigating road washouts. Based on a survey of road washout, 6 sites were selected for the study and the upland drainage areas of those drainage hotspots were identified. A cell-based watershed model, IMWEBs (Integrated Modelling for watershed Evaluation of Beneficial Management Practices), was applied to characterize landscape processes and simulate flow peak reductions in 6 road washout sites from restoring riparian buffers and wetlands in upland drainage areas. The IMWEBs results showed that the peak flow reduction ranged from 3.0% to 24.7% across these sites. The study also estimated economic costs of riparian and wetland restoration based on data from previous similar projects and the results showed that the natural infrastructure costs for flow peak reductions at the 6 road washout sites ranged from $477 to $10,586/yr across these sites. The study found that the costs for implementing natural infrastructure are comparable to the costs for implementing grey infrastructure such as installing larger culverts. Furthermore, the modelling and analysis results showed that there existed spatial variations of economic costs and flow reduction benefits of riparian and wetland restoration. There also existed spatial trade-offs between economic costs and flow reduction benefits of riparian and wetland restoration, ranging from $136/yr to $503/yr for 1% of flow reduction across these sites. These cost effectiveness analysis results have the potential to be used to identify priority locations for implementing natural infrastructures for mitigating road washouts in Parkland County.