Flood Susceptibility Analysis on Hexagonal Grid Meshes

In southern New Brunswick, Canada, seasonal flooding takes place around St. John River. According to the report by the  Government of New Brunswick , flooding is largely caused by meteorological conditions: thick snowpack remains for a long period each year until the spring thaw when snowmelt runoff flows into rivers, which leads to local overflows. Additionally, the accumulations of rain and the rising temperatures can accelerate snowmelt and surface runoff. Flooding can cause serious damage and inconvenience for citizens' the daily life, and flood susceptibility prediction can help the government and citizens to make  response strategies .

Flood modeling has been studied by using a bunch of methods, such as  artificial neural networks ,  support vector machine , and  decision-tree-derived  machine learning models. In addition, hydrological modeling on hexagonal grid meshes has drawn attention among researchers. Previous studies illustrated several key advantages of hexagonal cells compared to traditional square cells in hydrological modeling. For example, hexagons eliminate the ambiguity of the cell neighborhood due to its  uniform adjacency , and therefore, various weighting schemes and assumptions can be avoided in the cells’ neighborhood. Consequently, hexagonal grids remove the  island effect  which is an obstacle when modeling watersheds in square grids with both direct and diagonal neighbors. Moreover,  Discrete Global Grid Systems (DGGS)  was increasingly adopted in integrating multi-sources data and solving real-world problems. For example, a DGGS was used to predict flooding risks by adjusting the  height above the nearest drainage model , although some of the hydrological variables were computed in a traditional GIS manner before being quantized in the DGGS.

This project aimed to model flood susceptibility in a hexagonal DGGS:

1. Quantization of sample points and predictor variables in DGGS.
2. Compute meteorology variables in DGGS.
3. Derive topographical and hydrological parameters in DGGS.
4. Train random forest models across multiple granularities and filter out important predictors.
5. Predict and visualize the flooding susceptibility over the full study area.

The study area is around 7705 km ² , located in southern New Brunswick, and covers the partial drainage basin of St. John River. A total of 2795 sample points were randomly distributed within the study area extent. Their wet and dry attributes, namely flooded and non-flooded, were extracted from the historic record of flood events provided by  Open Maps .

Four categories of 28 predictor variables were included in the modeling: geomorphology, hydrography, meteorology, and terrain-derived variables.

•  Land cover classes of Canada (lc) 
• Forest cover percentage (fcp)
• Impervious areas (ia)
• Normalized difference vegetation index (ndvi)
•  Soil (sol) 
•  Surficial geology (geo) 

• Flow direction (fldir)
• Flow accumulation (flacc)
•  Mean snow and ice (msi) 
•  Distance to the nearest major waterbodies or flow lines (nhn) 
•  Wetland (wl) 
• Stream power index (spi)
• Topographic wetness index (twi)

•  Total precipitation (precip) 
•  Average temperature (tavg) 
•  Days > 10mm rainfall (r10) 
•  Days > 25 mm rainfall (r25) 
•  Days with min temp < -10 degree (tm10) 
•  Days with snow depth > 50 cm (sd50) 
•  Number of Spring days with min temp > 0 degree (spr) 
•  Total snow (ts) 

•  Digital Terrain Model (dtm) 
• Aspect (asp)
• Slope (slp)
• Curvature (curv)
• Roughness (rgh)
• Terrain roughness index (tri)
• Topographic position index (tpi)

The DGGS configuration used in this research was the  Icosahedral Snyder Equal Area Aperture 3 Hexagonal Grid (ISEA3H) . The DGGS gird was orientated with the latitude of the pole (λ) = 37.6895°, longitude of the pole (φ) = -51.6218°, and azimuth (α) = -72.6482°. Resolution levels 19, 21, and 23 were used in modeling, where the cell sizes were 43885.62, 4876.18, and 541.80 m ²  in area, respectively.

Sample points were quantized in the ISEA3H DGGS by converting their longitude-latitude coordinates to cell addresses of nearest cell centroids at levels 19, 21, and 23, using R library  dggridR . The quantization process included a series of interpolations over cell centroid locations of sample points. Specifically, predictors with factor and numeric values were sampled by nearest neighbor interpolation and bilinear interpolation, respectively. In terms of distance to the nearest major waterbodies or flow lines (nhn), vectors representing major waterbodies and flow lines were firstly rasterized and quantized in DGGS, and hexagonal ring numbers between each quantized sample point to the nearest waterbodies or flow lines were computed as nhn values.

Climate stations with normal climate data were determined with a 500 m buffer around the study area. Each meteorology variable value of sample points was calculated by Inverse Distance Weighted (IDW) interpolation, where the hexagonal ring numbers between quantized sample points and climate stations represented the distance.

Topographical predictors were derived from the quantized Digital Terrain Model (DTM). Predictors slope (slp) and aspect (asp) were calculated by the  Finite-Difference Algorithm . Predictor roughness (rgh) was the absolute difference between maximum and minimum elevations within the neighborhood (i.e., six directly connected cells). Predictor curvature (curv) was calculated as a  standard curvature  combining both profile and planform curvatures.

Before calculating hydrological parameters, depressions and flat areas were removed following the method proposed by  Barnes, et al. (2014) . Predictor flow direction (fldir) was computed following the  D6 algorithm  where the flow of a center cell was always routed to the neighboring cell with the lowest elevation, and upslope area, or flow accumulation (flacc), was then computed based on the flow routing grid.

Random forest was used to model and predict the flood susceptibility, where 70% and 30% of sample points were randomly selected for model training and testing, respectively. R library  VSURF  was used to rank variable importance using training data and predict the flood extent over the full area. Specifically, the important variables were selected by Interpretation Step in VSURF, and variables used to predict flooding were filtered by Prediction Step in VSURF. Model performances were evaluated by testing data with three indicators using  ROCR  library: accuracy (ACC), F-score, and area under the ROC curve (AUC).

Predictors used in prediction, ranked by variable importance, determined by library VSURF. Abbreviations are referred to the text above.

Level 19

• dtm, ts, precip, nhn, lc, rgh, sd50, tri, msi
• ACC = 0.920, AUC = 0.920, F-score = 0.917

Level 21

• dtm, ts, nhn, precip, geo, sd50, slp, r10
• ACC = 0.926, AUC = 0.925, F-score = 0.922

Level 23

• dtm, ts, nhn, precip, lc, rgh
• ACC = 0.942, AUC = 0.942, F-score = 0.938

Results showed that dtm was the most important variable, generally followed by hydro-geomorphological variables nhn, lc, and geo. Meteorology variables, precip and ts in particular, were listed as important predictors when being added to the models. It meant that total precipitation and snow had a strong impact on the occurrence of flooding. Such impact was slightly lower than that of elevations and distances to waterbodies, while higher than the other tested hydro-geomorphological variables.

Models performed well according to three evaluation indicators, where ACC, AUC, and F-score were higher than 0.9 across all resolution levels. Generally, models had better performance at finer resolutions. ACC, AUC, and F-score of HG8M model at level 23 were 0.942, 0.942, and 0.938, respectively.

Cell-based flooding events were predicted and visualized at three resolution levels. Although there were slight differences in the visualized flooding extent in various scenarios, predicted flooding sites were clustered around the St. John River and its branches.

1. This project modeled flood susceptibility in hexagonal DGGS.
2. DGGS helped to integrate multi-source data and conduct and visualize cell-based predictions.
3. DTM was the most important predictor variable.
4. Meteorology variables showed high importance.
5. Model performance was generally better at finer resolutions.
6. Flood susceptibility was predicted and visualized in a cell-based fashion.

The major challenge in this project was the amount of computation at fine resolutions. This was especially reflected in the production of flow direction and accumulation. One of the future directions is to parallelize these processes by decomposing the dataset into regular tiles or irregular sub-basins, solving each of them independently, and combining adjacent tiles or sub-basins by joining their edges. In addition, more machine learning methods, such as support vector machine, gradient boosting machine, and neural network can be applied and compared.

All data used were obtained from Canada's Open Government Portal, under the Open Government License - Canada.

This project was supervised by Dr. E. Stefanakis and Dr. H. McGrath. I acknowledge the funding from the NSERC Discovery Grant program and NSERC CREATE DOTS program.

• Barnes, R.; Lehman, C.; Mulla, D. Priority-flood: An optimal depression-filling and watershed-labeling algorithm for digital elevation models. Comput. Geosci. 2014, 62, 117-127.
• Chaudhuri, C.; Gray, A.; Robertson, C. InundatEd-v1.0: a height above nearest drainage (HAND)-based flood risk modeling system using a discrete global grid system. Geosci. Model Dev. 2021, 14, 3295-331.
• Hojati, M.; Robertson, C. Integrating cellular automata and discrete global grid systems: a case study into wildfire modelling. In Proceedings of the 23rd AGILE Conference on Geographic Information Science.
• Li, M.; McGrath, H.; Stefanakis, E. Integration of heterogeneous terrain data into Discrete Global Grid Systems. CaGIS 2021, 48, 546-564.
• Liao, C.; Zhou, T.; Xu, D.; Barnes, R.; Bisht, G.; Li, H.-Y.; Tan, Z.; Tesfa, T.; Duan, Z.; Engwirda, D.; et al. Advances in hexagon mesh-based flow direction modeling. Adv. Water Resour. 2022, 160.
• Robertson, C.; Chaudhuri, C.; Hojati, M.; Roberts, S.A. An integrated environmental analytics system (IDEAS) based on a DGGS. ISPRS J. Photogramm. Remote Sens. 2020, 162, 214-228.
• Wright, J.W. Regular hierarchical surface models: a conceptual model of scale variation in a GIS and its application to hydrological geomorphometry. Doctor of Philosophy, University of Otago, Dunedin, Otago, New Zealand, August 2017.

I am a Ph.D. candidate in the Department of Geomatics Engineering at the University of Calgary. I received my MSc in Forestry from the University of New Brunswick where I worked on the spatial analysis applied to the field of forest ecology. My current research interests are in the usability of Discrete Global Grid Systems, especially in the digital terrain data management and geospatial operation development in Discrete Global Grids.

 Personal Webpage  |  Research Gate  |  LinkedIn  |  GitHub  |  Email 

 _{This story map was submitted in partial fulfillment of the ESRI Young Scholars Award 2022 application.}