Tsunamis, primarily triggered by earthquakes, pose critical threats to coastal populations due to their rapid onset and limited evacuation time. Two main protective actions
exist: sheltering in place, which requires substantial retrofitting investments, and evacuation, which is often hindered by congestion, mixed travel modes, and tight inundation
times. Given pedestrians’ slower movement and restricted evacuation opportunities, vertical shelters within inundation zones provide essential refuge, while distant horizontal shelters outside these areas are often perceived as safer. As both shelter types complement each
other, they must be jointly planned. Evacuation effectiveness is further complicated by
noncompliance behavior, which, if neglected, can lead to inefficient retrofitting, capacity
overload, and even failure of evacuation strategies. This study introduces a time-expanded
multimodal evacuation optimization model that integrates shelter-in-place retrofitting, vertical shelter location, shelter assignment, and routing under tight inundation time, arc and
shelter capacities, and compliance likelihoods derived from a survey. The objective is to
minimize the time-dependent risk faced by evacuees at home, en route, and at shelters
while addressing unsatisfied demand. The model is solved through an integrated Benders Decomposition–Column Generation framework, where Column Generation generates
evacuation routes for compliant evacuees and shortest-time paths for noncompliant ones.
Convergence is enhanced through decomposition for parallel processing, network simplification, Maximum Density Cuts, and warm-start strategies. The model is applied to
Büyükçekmece, a densely populated and highly vulnerable district of Istanbul. Computational results highlight the necessity of incorporating compliance behavior, retrofitting,
and vertical shelters into evacuation planning, leading to substantial reductions in overall
risk and significant improvements in life safety and evacuation efficiency.