Towards a Transferable UAV-Based Framework for River Hydromorphological Characterization

Abstract

The multiple protocols that have been developed to characterize river hydromorphology, partly in response to legislative drivers such as the European Union Water Framework Directive (EU WFD), make the comparison of results obtained in different countries challenging. Recent studies have analyzed the comparability of existing methods, with remote sensing based approaches being proposed as a potential means of harmonizing hydromorphological characterization protocols. However, the resolution achieved by remote sensing products may not be sufficient to assess some of the key hydromorphological features that are required to allow an accurate characterization. Methodologies based on high resolution aerial photography taken from Unmanned Aerial Vehicles (UAVs) have been proposed by several authors as potential approaches to overcome these limitations. Here, we explore the applicability of an existing UAV based framework for hydromorphological characterization to three different fluvial settings representing some of the distinct ecoregions defined by the WFD geographical intercalibration groups (GIGs). The framework is based on the automated recognition of hydromorphological features via tested and validated Artificial Neural Networks (ANNs). Results show that the framework is transferable to the Central-Baltic and Mediterranean GIGs with accuracies in feature identification above 70%. Accuracies of 50% are achieved when the framework is implemented in the Very Large Rivers GIG. The framework successfully identified vegetation, deep water, shallow water, riffles, side bars and shadows for the majority of the reaches. However, further algorithm development is required to ensure a wider range of features (e.g., chutes, structures and erosion) are accurately identified. This study also highlights the need to develop an objective and fit for purpose hydromorphological characterization framework to be adopted within all EU member states to facilitate comparison of results.

Keywords: artificial neural network; hydromorphology; intercalibration; photogrammetry; unmanned aerial vehicle; water framework directive.

Figure 1

Schematic diagram showing the location of the selected study sites within each Geographical Intercalibration Group (GIG) and detailed imagery of the selected reaches. The maps of Spain and UK show the delineation of the main river basins with those basins containing the study sites highlighted in red.

Figure 2

Example of classification outputs obtained for each Geographical Intercalibration Group (GIG). From left to right, orthoimage, 2 m × 2 m ground truth grid and classified outputs from the Artificial Neural Network (ANN); (a–c) Outputs for the Central-Baltic GIG reach; (d–f) Outputs for the Mediterranean GIG reach; (g–i) Outputs for the Very Large Rivers GIG reach.

Figure 3

Types of Ground Control Points (GCPs) and Unmanned Aerial Vehicles (UAVs) used to collect the imagery at each reach. (a) 1 m × 1 m Squared GCP used in the Central-Baltic reach; (b) 0.30 m diameter GCP used in the Mediterranean and Very Large Rivers reaches; (c) IRIS9+ UAV (3DR, Berkeley, CA, USA) used at the Mediterranean reach; (d) Falcon 8 Trinity (ASCTEC, Krailling, Germany) used at the Central-Baltic reach; (e) md4-1000 UAV (Microdrones, Inc., Kreuztal, Germany) used at the Very Large Rivers reach.

Figure 4

Workflow followed from imagery collection to multiple comparison analysis. ANN, GCP and GSD stand for artificial neural network, ground control point and ground sampling distance, respectively.

Figure 5

Number of points of the 2 m × 2 m ground truth grid allocated to each feature for each of the Geographical Intercalibration Groups (GIGs).