Skip to main content Skip to local navigation

Predicting riverbeds in Nigeria using machine learning algorithms and ‘eyes in the air’

Predicting riverbeds in Nigeria using machine learning algorithms and ‘eyes in the air’

Adeyemi Olusola

by Adeyemi Olusola

Very few studies have been successful in combining ‘eyes in the air’ (remote sensing) with river characteristics to predict riverbed types (Channel Unit Types) within a river channel.

This study on River sensing: Inclusion of red band to predict reach-scale types using machine learning algorithms aims to understand the role of satellite sensors in predicting riverbeds using machine learning algorithms (MLAs). To achieve this aim, our study sought to answer the following questions:

● To what extent can variations in river characteristics help in predicting riverbed types using MLAs?

● Are there differences or similarities in generic classifications (by Rosgen) and statistical classification of riverbed types within the river channel?

● How effectively can MLAs aid in riverbed characterization?

Figure 1: The extent of the Ogun River Basin showing a section of the Upper Basin in detail.

Our study was able to predict riverbeds within the Upper Ogun River Basin in Southwest Nigeria by fusing remotely sensed data with river features utilizing MLAs (random forest, support vector machines, multiple adaptive regression splines, extreme gradient boosting, and adaptive boosting).

The Ogun River Basin (ORB) of Southwestern Nigeria is divided into the upper and lower ORB. ORB is located between latitudes 6°26ʹ and 9°10ʹN, and between longitudes 2°28ʹ and 4°8ʹE. The land area is about 23 000 km2 (Fig. 1).  The Oyan and Ofiki river systems make up the majority of the Upper Ogun River Basin (UORB), where this study was done. Its extreme points are about 200 km long and 140 km wide.

Both secondary and empirical data were used in the study (Fig. 2). Examples of secondary sources include topographical maps, geological maps, and digital elevation models. Primary sources are the characteristics of the river, which are identified and quantified in the field. We took 83 river lengths and 249 cross-sections from six third-order basins (Fig. 1). In addition, remotely sensed data (Landsat 8 and Sentinel-1) were collected and mosaiced using the Google Earth Engine platform.

Results

Figure 2: Methodology flowchart showing steps and processes for the cluster, control identification and prediction.

We used random forest-recursive feature elimination to find the most important variable(s) in identifying different types of riverbeds in order to achieve the purpose of this study. The top five significant factors (accuracy: 0.79 0.14; kappa: 0.39) that might help identify between different riverbed types are the dimensionless stream power, slope, breadth, wetted perimeter, and Band 4. In general, the mapping and classification of riverbeds using remote sensing holds a lot of promise. Olusola et al. stated in this study that statistical classification, which had already discovered distinctive bedforms local to the region of interest, was preferable to the descriptive categorization of river bedforms.

Conclusions

The site-specific riverbed statistical classification based on field measurements and observations once again demonstrates the shortcomings of broad classification methods in describing local observations. The current global classifications of riverbeds serve as a baseline for field mapping, but it is necessary to interpret observations on a merit-by-merit basis in order to carefully uncover site-specific types that could greatly help in solving local river problems and aid watershed management.

Although the Rosgen framework and the underlying statistical classification were found to be in agreement for some of the riverbed types, the observed differences nonetheless highlight the need for an objective approach to riverbed classification. According to the current situation, statistical classification can assist us in understanding how rivers respond to disturbances such as changing landuse, climate, and population; yet, incorrect and unprepared management might result in undesirable stream restoration actions.

This research article is co-authored with Onafeso Olumide, Olutoyin Adeola Fashae and Samuel Adelabu. Full citation: Adeyemi Oludapo Olusola, Onafeso Olumide, Olutoyin Adeola Fashae & Samuel Adelabu (2022). River sensing: the inclusion of red band in predicting reach-scale types using machine learning algorithms, Hydrological Sciences Journal, 67:11, 1740-1754, DOI: 10.1080/02626667.2022.2098752. Photo credit for Ogun River: Royreal, Wikimedia Commons.

EUC Assistant Professor Adeyemi Olusola is a river catchment scientist with a strong focus on river dynamics and human impacts on river catchments, as well as extreme events and reach-scale classifications. He is a physical geographer with a special interest in fluvial geomorphology and ecohydrology as well as GIS/remote sensing. One of his contributions to humid-tropical geomorphology,  together with some colleagues (Late Emeritus Prof Adetoye Faniran, Prof Lawrence Jeje and Dr. Olutoyin Fashae) is an edited book on Landscapes and Landforms of Nigeria under Springer’s World Geomorphological series. The book is expected to be out by March 2023.