In September 2015, moisture from Typhoons 17 (Kilo) and 18 (Etau) converged over northern Honshu Island, Japan. Intense rain bands stalled between the cyclones over the central mountain range producing exceptional rainfall up to 660 mm over 5 days. On the Kinugawa River, four flood control dams spilled and sent record flows down the steep river. The river rose quickly (10 meters over 24 hours) and prompted warnings of possible flooding. Early on 10 September a low point began to overtop in the municipality of Joso-shi and erode the natural high ground sending floodwater through rural neighborhoods and many hectares of rice fields. The river continued to rise and overtopped other low points in the left bank levee. Shortly before 13:00 hours, a 4 m tall levee breached directly adjacent to several homes, releasing a powerful flood wave capable of toppling homes and floating away cars. The breach flow combined with upstream flow from a minor breach and overbank flooding, and began to pool near the downtown area of Joso, trapping hundreds of people in their homes. The disaster resulted in two fatalities. National and local government agencies responded by performing thousands of rescues by air, water, and even carrying people to safety on their backs.
The following provides a short description of the data sources used for validation. The models for this validation effort was created by USACE and benefited from data from social media augmented official sources to create a complete and accurate data set.
The following excerpt was taken from Risher et al (2017): River hydraulic modeling started with pre-flood LIDAR terrain data for the Kinugawa and flooded area. […]No tributaries were in the modeled reach. Modeled levees use the lateral structure option to allow overtopping and breach. The large canal gate near downtown Joso allows flow to exit the flooded area once river levels begin to recede. A crucial missing component was the inflow hydrograph that the team had to derive from observed stage data at the Kamaniwa gage, at 27.34 km (upstream of the overtopping area). MLIT provided a modeled stage hydrograph at the breach location (21.0 km) to calibrate the inflow. […] Many observations were made of the breach through the course of the event with their times noted and breach width estimated.
Calibration of the hydraulic model used data from a variety of sources. Several time-stamped aerial photos and Japan’s Geospatial Information Authority (GSI) maps outlined the flood limits. Geo-located photos of high water marks provided over seventy-five maximum depth estimates. Time-specific depths came from photos, video, and eyewitness descriptions. Calibration focused on arrival time and maximum depth - the two criteria most important to evacuation and lifeloss modeling. Velocities were also spot checked where structures failed. One early problem with the river hydraulics model was that water was moving too fast through the floodplain trapping more people than expected. To match observed conditions the team added several road and railway embankments causing water to pool and overtop them sequentially before moving on.
Structure Inventory and Road Network
LifeSim requires an inventory of all possibly affected structures as the starting point for population. A structure inventory was created by manually placing a point on each building observed in recent aerial imagery. The limits of the survey were the Kinugawa River on the west, the Kokaigawa River on the east, and a 100 m buffer beyond the observed flooding limit to the North and South. Each building has a damage category, occupancy type, foundation height, number of stories, and construction material. The structure locations were checked against OpenStreetMap while the other parameters were checked with Google Street View. Then the structures were populated based on their occupancy type for daytime and nighttime conditions. The population estimate for Joso started with 2010 adjusted census data at the neighborhood level [10, 11]. It was indexed based on 2015 population growth rates at the shi level and applied uniformly across all census blocks within the shi. There was a small decrease in population from 2010 to 2015. The final population estimate was 28,052 in Joso-shi and 33,219 total including small parts of adjacent shi that were flooded. The model population is split into groups over and under age 65 based on the 2010 shi-level ratios (roughly 21-23% elderly).
The breach zones got special attention where deep and fast water can destroy buildings. These areas accounted for a significant portion of the lifeloss. Using video and other evidence of the rescue effort, the homes in the breach area were populated with the correct number of people at home. The structures were scrutinized more closely for foundation heights, construction materials, number of stories, etc. so that the proper depth and structural stability curve would be applied. In LifeSim, if a building collapses all the people inside are subject to the highest chance fatality rate (average of 91.45%) so it is important to define those buildings properly.
The road import tool in LifeSim automatically builds a road network from OpenStreetMap data, used under Open Database License. Each road segment has a traffic capacity rating based on its size and design (e.g. highway vs. street). The team checked the network for one-way streets, unnecessary segments, and traffic capacity to prevent too many people trapped on roads while evacuating. Any bridges or elevated roads are vertically offset to prevent them from flooding.
Emergency Planning Zones
In Japan as in the US, local municipalities give evacuation warnings while a national agency gives severe weather statements. Evacuation warnings come in the form of pre-evacuation information, evacuation recommendations, and evacuation orders. All three types of warnings were given at various times and places by local government officials based on their perception of the risk. […] For modeling purposes, it was assumed people outside of the warned zones may still have gotten a warning unintentionally or indirectly, but the warning dissemination will be slower. It was also assumed the recommendation at 04:00 was not as good as an order.
The PAI curve defines the time when evacuation begins. People evacuate by car following the road network from a structure to the nearest destination point. Flatter curves indicate populations are slower to evacuate after hearing a warning. Typically a maximum rate is set below 100% because some portion of the population will choose not to leave, or cannot leave. For this event, the max rate was set at 58.5% based on responses from the survey. This rate aligns with anecdotal evidence from Japan that evacuation rates are typically low, but there is no solid information about how many, how fast, or by what means people evacuated. The PAI curve is the most uncertain parameter in the consequence model due to the lack of information. There were reports of traffic congestion and thousands of people in shelters, but also thousands rescued by boat and helicopter, and people walking out through the flood. Rescue and sheltering are not accounted for in model evacuation.
Destinations were set at each bridge out of the leveed area and roads leading north and south out of the flooded area.
To simulate the staged warnings and mix of official and unofficial warnings, each EPZ was given its own warning time relative to breach initiation. Critically, the zone nearest to the breach zone was assumed to receive warning after breach; however, based on PAR were removed from the structures near the breach when available evidence suggested the homes were empty at the time of the breach.
Risher, Paul; Ackerman, Cameron; Morrill-Winter, Jesse; Fields, Woodrow; Needham, Jason (2017). Levee Breach Consequence Model Validated by Case Study in Joso, Japan
If you have any questions, please e-mail RMC.Software@usace.army.mil