66 | INTERPRAEVENT 2016 – Extended Abstracts
INTRODUCTION
In the research field of debris and bedload trans- port, the calculation and estimation of event-based debris yield is challenging. As debris floods or debris flows cause high losses to buildings and infrastruc- ture; torrents are managed by check dams and barriers. To improve estimations of event-based debris yield, considering also ecological and safety aspects, further research pathways are necessary.
Empirical formulas estimating the event-based debris and bedload yield, e.g. Zeller (1985), corre- lated the yield to catchment parameters (catchment area, channel slope etc.). However, Rickenmann (2014) highlighted for most events a considerable difference between the transported bedload and debris material of documented events and the calculated debris yield. There are two main reasons for these differences: In the empirical approaches, factors such as the catchment parameters, macro- roughness, unconsolidated material, precipitation conditions during the event and the proportion of vegetation, are not considered although, there are relationship between those parameters and trans- ported debris yield (Kienholz, 1998). A second weakness of the empirical approaches is the appli- cation of mathematical linear or exponential relations between the event-based debris yield and the parameters. To the authors‘ knowledge, there are only a few studies, which analyzes the non- linear multivariate relations between the transport- ed debris yield of events and the mentioned catch- ment parameters. Artificial neural networks offer this analysis option. Neural network is a soft computing approach, which is characterized by exact models for data with imprecisions and uncer- tainties (Zadeh, 1994).
OBJECTIVES
The aim of this study is to identify torrents with documented event-based bedload yield and to calculate the mentioned catchment parameters
(include precipitation conditions of the event) using GIS. In a second step, the relationships between the parameters and the event-based bedload yield will be analyzed with neural net- works.
DATA AND METHODS
A large database with different torrents and events is needed for the statistical analysis. The database of this study includes torrents of DB Solid (BAFU),) and the compilation of events of Gertsch (2009) and Rickenmann & Koschni (2010). For all tor- rents, the catchment parameters are calculated, including the above mentioned neglected param- eters. As base to determine these parameters the digital elevation model (DTM) swissAlti3D (swis- stopo), the vector model TLM3D (swisstopo) for land use information and the Rhires (meteosch- weiz) database for precipitation information is used.
For a systematically and normalized survey of the parameters a newly developed ‚R script‘ with a
‚QGIS‘ user interface is applied. The user-friendly interface reduces the steps for the user to integrate the DTM, specify the coordinates of the torrent fan and the date of the event. After determining the parameters and transported bedload, the database is analyzed with a neural network. In this study only the ‚feedforward-backpropagation network‘ is use.
The network consists of an ‚input-layer‘ (Fig. 1) to
Figure 1. Schematic diagram of feedforward-backpropagation network.
(based on Rey & Wender, 2011)
IP_2016_EA164
Improving debris yield estimation with neural network approach
Jan Baumgartner, BSc1; Margreth Keiler, PD Dr.1
DATA ACQUISITION AND MODELLING (MONITORING, PROCESSES, TECHNOLOGIES, MODELS)
INTERPRAEVENT 2016 – Extended Abstracts | 67
feed the independent variables (e.g. catchment parameters). In the ‚hidden-layer‘ these input values are further processed to an ‚output‘. The values will be multiplied with the weights between the layers. The ‚output‘ represents the independent variable (e.g. transported bedload). Using a large dataset with known output values, the difference between the modelled output and the documented output is calculated. To adjust the weights between the layers (training), an algorithm is applied. After the training, a phase follows where the fitted model is tested with independent datasets. The advantage of this approach is, that a ‚Feedforward-backpropa- gation network‘ with several ‚hidden-layers‘ can approximate any mathematical function (Kriesel, 2007:127). This is a promising approach because several several studies exist (e.g Sasal et al., 2009), which present the applicability to estimate for example bedload transport rate with channel parameters reaching a high accuracy using neural networks.
RESULTS AND DISCUSSION
The results of the new-developed GIS-tool com- prise now 64 parameters for 110 different torrent catchments. At least one field estimation of the volume of debris for an extreme-event exist for all of these 110 torrents. This generated information is the training and testing data for the neural net- work. Actually, a technique for the determination of the selection of different catchment parameters as input for the neural network is under develop- ment.
LITERATURE:
- Gertsch E. (2009). Geschiebelieferung alpiner Wildbachsysteme bei Grossereignissen - Ereignisan- alysen und Entwicklung eines Abschätzverfahrens.
Universität Bern.
- Kienholz H., Keller H.M., Ammann W., Wein- gartner R., Germann P.F., Hegg C., Mani P., Ricken- mann D. (1998). Zur Senisitivität von Wildbachsys- temen. Schlussbericht NFP 31. Forschungsbericht, vdf Hochschulverlag AG an der ETH Zürich, Zürich - Kriesel D. (2007). Ein kleiner Überblick über Neuronale Netze, erhältlich auf http://www.
dkriesel.com
- Rickenmann D. (2014). Methoden zur quantita- tiven Beurteilung von Gerinneprozessen in Wild- bächen. WSL Bericht 9: 105 S.
- Rickenmann D., Koschni A. (2010). Sediment loads due to fluvial transport and debris flows during the 2005 flood events in Switzerland.
Hydrological Processes, 24(8), 993-1007.
- Sasal M., Kashyap S., Rennie C.D., Nistor I.
(2009). Artificial neural network for bedload estimation in alluvial rivers. Journal of Hydraulic Research, 47(2), 223-232.
- Zadeh L.A. (1994). Fuzzy Logic, Neural Networks, and Soft Computing. Commun. ACM, 37(3), 77-84.
- Zeller J. (1985). Feststoffmessung in kleinen Gebirgseinzugsgebieten. Wasser, Energie, Luft, 77, 246-251.
KEYWORDS
neural networks, debris flow, bedload yield, torrent, assessment/estimation
1 University of Bern, Bern, SWITZERLAND, jan.baumgartner@students.unibe.ch