A Practical Framework for Rapid Earthquake Damage Estimation through Multi-Vehicle Coordination and Inter-Agency Collaboration: Integrating Genetic Algorithms with RGB-Based Image Analysis

A Practical Framework for Rapid Earthquake Damage Estimation through Multi-Vehicle Coordination and Inter-Agency Collaboration: Integrating Genetic Algorithms with RGB-Based Image Analysis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Practical Framework for Rapid Earthquake Damage Estimation through Multi-Vehicle Coordination and Inter-Agency Collaboration: Integrating Genetic Algorithms with RGB-Based Image Analysis Haruhiro Shiraishi, Yuichiro Usuda This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7390282/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 18 Nov, 2025 Read the published version in Natural Hazards Research → Version 1 posted You are reading this latest preprint version Abstract Quickly and accurately evaluating building damage after a major earthquake is essential for effective emergency response. We propose a practical framework that integrates camera-equipped ground vehicles deployed by multiple agencies (such as the Self-Defense Forces, police, and fire services) with a centralized command system to perform real-time post-earthquake damage mapping. The system combines an RGB-based damage detection technique (gNCDI), which generalises the simple Redness Index (RI) originally developed for vegetation analysis, with a Genetic Algorithm (GA) to optimise the patrol routes of multiple vehicles. Using colour-based inference, collapsed buildings are rapidly identified from ground-level images by detecting the spectral signatures of exposed timber and soil debris, while the GA efficiently allocates routes to each vehicle to maximise coverage and minimise response time. A cloud-based architecture standardises and shares geotagged damage reports in real time using a JSON format across all responding agencies. We present the system design, implementation details, and evaluation protocol based on a disaster scenario simulation for the Noto Peninsula region in Japan. In our evaluation, the proposed approach achieved a high overall classification accuracy (F1 score ≈ 0.86), detecting 90% of collapsed buildings with only ~ 18% false alarms. At the same time, the cooperative vehicle-routing strategy significantly improved survey efficiency, shortening total mission completion time by around 25% compared to a greedy baseline. Furthermore, we discuss practical issues including the speed and resolution advantages over traditional satellite or aerial assessments, data privacy considerations, false detections, and the need for human verification of results. Overall, this study demonstrates a feasible multi-vehicle, multi-agency approach for rapid earthquake damage estimation aimed at accelerating life-saving rescue operations and optimising resource allocation. Computer Architecture and Engineering Rapid Earthquake Damage Estimation Ground-Level RGB Image Analysis (gNCDI) Cooperative Multi-Vehicle Routing (Genetic Algorithm) Inter-Agency Coordination (ICS) Real-Time Disaster Mapping Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction After a strong earthquake, emergency responders need to obtain a rapid and accurate overview of building damage. Information on the locations of collapsed buildings and the severity of damage is crucial for guiding search-and-rescue efforts and efficient resource distribution. Traditionally, remote sensing methods using satellite imagery or aerial photographs have been employed for damage assessment [1]. However, these approaches suffer from temporal delays and practical constraints. For example, satellites have limited imaging opportunities and are easily hindered by cloud cover, and although helicopters or drones can capture high-resolution data, they typically cannot cover broad areas quickly and lack immediacy. Advanced image analysis techniques (such as detailed 3D map reconstruction [2] or complex anomaly detection algorithms [3]) are computationally intensive and often unsuitable for immediate post-disaster response. As a result, in practice ground visual surveys by municipal engineers, fire services, and police are a primary information source[4]—but these are labour-intensive and tend to be siloed by agency, leading to fragmented information. To bridge this gap, we propose a framework that combines multiple ground vehicles with a centralized system to automatically generate a reliable damage map within a few hours of an earthquake. This idea draws inspiration from recent technological trends: during the 2024 Noto Peninsula earthquake[5], for instance, vehicles equipped with 360° street-level cameras[6] were deployed to drive around the affected area and collect panoramic imagery. Such “drive-by” imaging can acquire close-range, high-resolution photos far more rapidly than foot surveys (Fig. 1 ). Moreover, coordinating vehicles from multiple agencies in a collaborative manner could potentially cover a wide area in a short time. Indeed, prior research also suggests the usefulness of ground mobile imaging. For example, Mizui and Fujiwara (2021) analysed vehicle dashboard camera footage to estimate disaster waste volume, demonstrating that a quick survey using in-vehicle video after the Hitoyoshi City floods enabled an accurate assessment of debris to be removed [7]. This indicates that rich information on damage extent can be extracted from footage captured by moving vehicles, which motivates our use of ground-level image data in this work. Advances in automated image analysis technology further support this research. Traditional computer-vision methods [8] (e.g. colour indices, texture analysis, edge detection) have been used to classify building damage, but in recent years deep learning approaches like Convolutional Neural Networks (CNNs) have become predominant [9]. However, deep learning requires large amounts of training data and computational resources, and the models act as “black boxes” with opaque decision logic. In the immediate aftermath of a disaster, a method that, even if simple, is rapid, explainable, and easily adjustable in the field is preferable. We focus on a simple colour-based index known as the Redness Index (RI)[10][11][12]. RI is a normalized difference computed from red and green colour intensities, defined by the following formula (1): $$\:\text{R}\text{I}=\frac{R-G}{\hspace{0.17em}R+G\hspace{0.17em}}$$ 1 Originally proposed as an index to correct vegetation measurements in arid regions[10][11], RI is particularly responsive to materials with strong red reflectance and weak green reflectance. We hypothesise that this spectral characteristic can serve as an indicator of severe structural damage: when a building collapses, structural timber becomes exposed and dust from roof tiles or earth plaster tends to increase the relative red reflectance. Because it requires no extensive training data and can be implemented as simple per-pixel calculations with thresholding, RI is well-suited to real-time processing in the field. Shiraishi and Usuda (2025) proposed the concept of using RI (a Normalized Red–Green Index) for building damage extraction and showed that RI could detect nearly all completely-collapsed buildings (damage grade D5 on the European Macroseismic Scale EMS-98) in the Noto Peninsula earthquake data[12]. In their preliminary validation, 14 out of 15 D5-level collapsed buildings were correctly detected by an RI threshold, and 14 out of 16 non-collapsed buildings were correctly identified as undamaged (most false alarms were intact buildings with red-colored roofs or walls)[12]. Building on the high recall and accuracy demonstrated by this simple colour index, our framework generalises RI into a generalized Normalized Chromatic Difference Index (gNCDI) (see Eq. ( 2 ) below) and places this at the core of a large-scale damage mapping approach. $$\:\text{g}\text{e}\text{n}\text{e}\text{r}\text{a}\text{l}\text{i}\text{z}\text{e}\text{d}\:\text{N}\text{o}\text{r}\text{m}\text{a}\text{l}\text{i}\text{z}\text{e}\text{d}\:\text{C}\text{h}\text{r}\text{o}\text{m}\text{a}\text{t}\text{i}\text{c}\:\text{D}\text{i}\text{f}\text{f}\text{e}\text{r}\text{e}\text{n}\text{c}\text{e}\:\text{I}\text{n}\text{d}\text{e}\text{x}\:\left(\text{g}\text{N}\text{C}\text{D}\text{I}\right)=$$ $$\:a\cdot\:\frac{R-G}{\hspace{0.17em}R+G\hspace{0.17em}}\hspace{0.33em}+\hspace{0.33em}b\cdot\:\frac{G-B}{\hspace{0.17em}G+B\hspace{0.17em}}\hspace{0.33em}+\hspace{0.33em}c\cdot\:\frac{B-R}{\hspace{0.17em}B+R\hspace{0.17em}}+d・\frac{R}{\hspace{0.17em}R+G+B\hspace{0.17em}}+e・\frac{G}{\hspace{0.17em}R+G+B\hspace{0.17em}}+f・\frac{B}{\hspace{0.17em}R+G+B\hspace{0.17em}}+\:g\text{・}\text{S}+h\text{・}\text{c}\text{o}\text{s}\left(2{\pi\:}\text{h}\right)+\text{i}\text{・}\text{s}\text{i}\text{n}\left(2{\pi\:}\text{h}\right)$$ 2 Here, RGB equals R, G, B in a linear intensity space. Let HSV be (H, S, V) with H in [0,1) and S in [0,1]. Another challenge is the coordinated route planning for multiple vehicles. In a large-scale disaster, various organisations (municipal authorities, fire brigades, police, Self-Defense Forces, etc.) each patrol their assigned areas to assess damage. Without sufficient coordination, however, survey efforts may overlap or miss certain areas. To enable efficient collaboration, we model the problem as a time-constrained multi-agent tour on a road network, and apply a Genetic Algorithm (GA) for optimising the routes of multiple vehicles. Vehicle routing problems (VRPs) and their variants have been widely studied in the logistics domain, and their application to disaster response has also been explored. For example, Taniguchi et al. (2000) investigated a probabilistic, time-varying travel time model accounting for uncertainty in disaster situations [13]. Kefi and Ghedira (2004) proposed a cooperative multi-agent solution for the VRP with time windows[14]. Zidi et al. (2013) applied a GA to disaster relief vehicle routing, dynamically incorporating new requests (rescue calls, etc.) and interruption events (e.g. road closures) to achieve a solution that maximised the number of affected people reached while minimising total travel cost [15]. Our approach follows the same principle, performing a multi-objective optimisation that maximises survey coverage (number of damaged buildings detected) while minimising the overall completion time for all vehicles. With regard to multi-agency information sharing, a unified command system such as the Incident Command System (ICS) is internationally adopted. Our proposed centralized coordination server aggregates and analyzes data from all vehicles regardless of their parent organisation, and dispatches task assignments or route changes to each vehicle in real time. This centralized control enables all agencies to operate with a common, up-to-date damage map and progress status, while maintaining agility in the field. In summary, the framework presented in this research is a novel rapid damage assessment system that combines ground-level image analysis with cooperative vehicle routing. The main contributions of this work are as follows: Methodology: We propose an earthquake damage mapping approach that integrates a simple colour-based damage detection technique with a multi-vehicle route optimisation strategy. System Design: We design a cloud-based, real-time information-sharing system for inter-agency collaboration, using a standardised JSON data format for geospatial damage reporting. Validation: We verify the effectiveness of the proposed approach through simulation of a scenario in Noto Peninsula, and conduct a performance comparison with conventional methods (e.g. a greedy heuristic). Practical Considerations: We discuss issues for real-world deployment, including ethics, privacy, and data reliability. The remainder of this paper is organized as follows: Section 2 describes the proposed framework in detail, including the image-based damage detection and cooperative vehicle routing methods. Section 3 outlines the implementation and dataset used in our simulation. Section 4 presents the evaluation results, and Section 5 provides discussion of advantages, limitations and practical implications. Finally, Section 6 offers concluding remarks. 2. Methodology In this section, we explain the core techniques of the proposed framework. First, we describe the image-based damage detection method using the gNCDI index. Next, we explain the road network data used for route planning and how hazardous areas are handled. Finally, we detail the design of the cooperative GA (Genetic Algorithm) — including the chromosome representation, evolutionary operators, and the route optimisation objective function. 2.1 Damage Detection Using gNCDI and GA To enable real-time damage interpretation on frames transmitted from vehicles in disaster areas, we combine the generalized Normalized Chromatic Difference Index (gNCDI) with a Genetic Algorithm (GA). gNCDI generalizes the conventional Redness Index (RI). While RI essentially uses only a single “red–green difference,” our approach employs a 9-dimensional color feature vector. This increases sensitivity not only to reddish debris but also to cues that RI tends to miss, such as blue tarpaulins and fine cracks. 9-Dimensional Features and Definition of gNCDI Let RGB be R, G, B in a linear intensity space. Let HSV be (H, S, V) with H in [0,1) and S in [0,1]. Let eps = 1e-6 be a small constant to avoid division by zero. Define the 9-dimensional feature vector x as: x1 = (R - G) / (R + G + eps) x2 = (G - B) / (G + B + eps) x3 = (B - R) / (B + R + eps) x4 = R / (R + G + B + eps) x5 = G / (R + G + B + eps) x6 = B / (R + G + B + eps) x7 = S x8 = cos(2 pi H) x9 = sin(2 pi H) With coefficient vector w = [a, b, c, d, e, f, g, h, i] and bias b0, the gNCDI score is: gNCDI(x) = a x1 + b x2 + c x3 + d x4 + e x5 + f x6 + g x7 + h x8 + i*x9 + b0 The first three terms (opponent color differences) are robust to illumination changes and capture red–green, green–blue, and blue–red contrasts. The next three terms (RGB proportions) reflect color composition relative to overall brightness. The last three terms embed hue on the unit circle via cos and sin and include saturation. As a result, the model responds not only to “redness” (RI-like debris) but also to blue–red contrasts (e.g., blue tarps) and subtle chromatic irregularities associated with cracks. Learning with GA (F1 Maximization) We use gNCDI(x) as a damage score and classify a pixel (or image) as “damaged” if it exceeds a threshold tau. During training, the coefficients w and bias b0 are jointly optimized by GA, and for each candidate solution we internally search the threshold that maximizes the F1-score based on the score distribution. Procedure: (i) initialize a random population; (ii) evaluate each individual by computing precision and recall, then its F1 as fitness; (iii) keep elites via elitist selection; (iv) generate offspring by crossover (real-valued recombination of weights and bias); (v) apply mutation (small perturbations) to maintain diversity. This yields (w, b0, tau_star) tailored to local conditions even with a small calibration set. Deployment (Real-Time Inference) At inference, compute x for all pixels in each frame, evaluate gNCDI(x) with the optimized w and b0, and extract pixels with scores > = tau_star as damage-indicative pixels. If high-scoring pixels form clusters exceeding a set area, flag the location as a damage candidate; voting across consecutive frames further reduces false alarms. The method remains lightweight and easy to implement like RI-based approaches, while the 9-D formulation captures richer damage cues, improving detection for cases difficult for RI alone (blue tarps, fine cracks). 2.2 Road Network Data and Hazardous Area Handling Efficient touring by multiple vehicles critically depends on leveraging road infrastructure data. In this study, the road network of the target region is provided in GeoJSON format, which we convert into a graph data structure for route planning. Specifically, each road segment (LineString in GeoJSON) is parsed into nodes and edges: every intersection or endpoint becomes a node, and connecting road segments are represented as edges. The length of each edge is calculated from the haversine distance of the corresponding road segment (the great-circle distance based on latitude/longitude), as an approximation of driving distance. We normalise node coordinates by rounding latitude/longitude to a consistent precision and merge any duplicate nodes as a preprocessing step to ensure a clean, connected graph. Since road blockages are likely to occur immediately after an earthquake, we incorporate known hazardous areas into the route planning to exclude those segments. The framework allows optional input of hazard polygons (areas of concern such as potential sinkholes, structural collapses, or severely flooded zones). For each provided hazard polygon, the system checks each edge of the road graph to determine if it geometrically intersects or lies within the polygon. Any such edge is marked as impassable and removed from the graph. This ensures that the routing algorithm will automatically avoid paths through dangerous areas. After removing edges, any node that becomes isolated (no longer reachable from any other node) is removed from the set of targets to maintain overall graph connectivity. We fix the vehicle travel speed at ~ 40 km/h (typical urban driving speed) for the simulation, converting graph distances to travel times by this speed. The primary route cost calculations use this travel time as the weight. In real operations, one should account for traffic obstructions or congestion, but for simplicity in our simulation we assume a constant speed (it is possible to pre-assign per-road penalties based on expected travel difficulty, if needed). We assume one depot (base) location for each agency, i.e. each of the 9 vehicles starts and ends at its designated depot. Depot coordinates can be provided as a set of GeoJSON Points; if no depots are specified, we randomly placed nine depots near the area boundary in our simulation. The points of interest (targets) represent the locations of buildings or areas to be inspected. We set approximately 70 target points uniformly sampled from the road network nodes, which can be thought of as representing the buildings to survey in the region. Furthermore, we marked about 20 of these as points where “building collapse occurred” in the scenario. The collapse points were selected at random (uniformly distributed across the area) in this study, but in practice they would be determined dynamically by the gNCDI analysis module in real time. In other words, the gNCDI processing module would flag locations with “possible” values as “suspected collapse,” which would then be shared with the route planning module. In our simulation, we simply fixed 20 points as collapsed at the start, but the framework is designed to operate under the assumption that the set of collapse flags is updated dynamically during the mission. Once the road graph and the depots/targets are prepared, we pre-compute a distance table to be used for route planning. We calculate the shortest-path distance on the graph between every pair of POIs (Points of Interest: roughly 70 targets + 9 depots, ~ 79 points in total) using Dijkstra’s algorithm, and convert those distances to travel times to form a time cost matrix \(\:T\) . Any pair that is unreachable (for example, if two points are on separate islands with no connecting roads) is assigned an infinite distance. This preprocessing speeds up the cost evaluations and the optimisation procedure described next. 2.3 Cooperative GA: Chromosome Design and Objective Function For route optimisation, we adopt a cooperative Genetic Algorithm (Cooperative GA, abbreviated Coop-GA). In this approach, a single individual (solution) represents the route assignments for all vehicles, and evolutionary computation is used to search for improved solutions. Below, we describe the design of the GA method used in this study. Chromosome Representation: One individual encodes the set of tour routes for all vehicles. Specifically, it is represented as a tuple concatenating the ordered list of assigned targets for each of the 9 vehicles (e.g. an individual could be encoded as a tuple like ((…),(…), … ,(…)), where each component list corresponds to one vehicle’s target visitation sequence). Each route is a sequence of departing from the vehicle’s depot, visiting its assigned targets in order, and returning to the depot. However, the depot is not explicitly included in the chromosome; each vehicle’s gene sequence consists only of target IDs. During evaluation, the depot of each vehicle is automatically prepended and appended to that vehicle’s target list to form the complete circuit. Initial Population: Rather than purely random initialisation, the initial population of solutions is seeded with a heuristic solution to ensure both diversity and quality. First, we generate one heuristic route plan by assigning each target to the vehicle with the nearest depot, and ordering each vehicle’s targets roughly by nearest-neighbour (shortest incremental distance). This provides a plausible initial solution (we also generate the Greedy solution described later, and include it as well). We inject this heuristic solution (and the Greedy solution) into the initial population, while generating the remaining individuals with random task allocations and orders. This way, the initial population contains a few relatively good solutions while still maintaining broad variety. Fitness Evaluation (Objective Function): The GA’s evaluation function (fitness) is a composite cost metric for the set of routes across all vehicles. We define a weighted sum objective function that integrates three aims: (i) minimising the overall completion time, (ii) arriving at collapsed sites as early as possible, and (iii) balancing the workload among vehicles. Formally: $$\:J={w}_{1}\hspace{0.17em}{T}_{\text{m}\text{a}\text{x}}\hspace{0.33em}+\hspace{0.33em}{w}_{2}\sum\:_{i\in\:Collapsed}{t}_{i}\hspace{0.33em}+\hspace{0.33em}{w}_{3}\hspace{0.17em}\sigma\:$$ Here, \(\:J\) is the objective value for a given complete plan of routes for all vehicles. \(\:{T}_{\text{m}\text{a}\text{x}}\) is the maximum time taken by any vehicle to complete its route (i.e. the time when the last vehicle finishes, also known as the makespan). \(\:{t}_{i}\) is the arrival time at collapse-flagged target \(\:i\) (and the summation is the total of all such arrival times across identified collapse points). \(\:\sigma\:\) is the standard deviation of individual vehicle mission times (the time from each vehicle’s departure to its return to depot). The weight coefficients \(\:{w}_{1},{w}_{2},{w}_{3}\) are set according to the relative importance of these objectives. In our implementation, we assign weights such that reducing \(\:{T}_{\text{m}\text{a}\text{x}}\) is by far the top priority (i.e. \(\:{w}_{1}\) is orders of magnitude larger than the others), while also penalising delays in reaching collapsed sites and large disparities in vehicle workloads. In our standard setting, for example, we heavily weight \(\:{T}_{\text{m}\text{a}\text{x}}\) to ensure the fleet’s total completion time is minimised, and choose smaller but non-zero weights for the collapse arrival term and the time deviation term. In effect, this single scalar objective encodes a multi-objective optimisation that seeks to shorten the overall mission time (minimise \(\:{T}_{\text{m}\text{a}\text{x}}\) ), ensure prompt visits to collapsed-building locations (minimise the sum of collapse arrival times), and equalise the workload across vehicles (minimise the time standard deviation). As a result, the GA will assign higher fitness to solutions where no single vehicle is extremely delayed (due to the \(\:{T}_{\text{m}\text{a}\text{x}}\) term), collapsed sites are visited early in the routes (due to the second term), and each vehicle finishes in roughly the same time (due to the \(\:\sigma\:\) term). Selection, Crossover, and Mutation: The GA’s core operations are similar to their application in other combinatorial optimisations. In each generation, selection is applied based on objective value (using techniques like roulette-wheel or tournament selection) so that better individuals are more likely to be chosen for reproduction. For crossover, we adopt a vehicle-wise route segment crossover . That is, for each of the 9 vehicles, we randomly extract a portion of the route (a contiguous subsequence of targets) from parent A and parent B, and combine them to form each vehicle’s route in the child. In this way, a single crossover operation recombines the routes of all vehicles simultaneously. After crossover, the child solution may contain duplicate visits or unassigned targets; we apply a repair operator to remove any duplicate target assignments and reinsert any missing targets, thereby maintaining a valid complete solution. For mutation, we introduce small random modifications to an individual to promote diversity. We combine multiple mutation operators in our implementation【15】, including: Inter-vehicle relocation: Remove one task from the most heavily-loaded vehicle and insert it into the route of the least loaded vehicle. This adjusts the balance of task counts between vehicles, alleviating extreme disparities (directly reducing the standard deviation term in the objective). Inter-vehicle swap: Randomly pick two vehicles and swap one task between their routes. By transferring a relatively out-of-the-way task to another vehicle, this operator explores the possibility of reducing overall distance/time. Intra-route 2-opt shuffle: For each vehicle’s route, either swap the order of two visits that are very close to each other, or take a randomly selected segment of the route and reverse its order. This is analogous to the classic 2-opt method in the Traveling Salesman Problem, and serves as a local optimisation to shorten an individual vehicle’s route. After any mutation, a repair step is again performed to fix any inconsistencies (e.g. duplicate or missing tasks). By applying these operators (potentially multiple times per generation) in combination, the algorithm explores a wide search space while still guiding the population towards convergence. Local Search Enhancement: Once the GA solutions have converged to a reasonably good region, we further apply a local search to fine-tune and improve the solution. Specifically, we sequentially apply: (i) a 2-opt optimisation on each vehicle’s route to remove any unnecessary detours, (ii) a relocation of one task from the slowest vehicle to the fastest vehicle to directly reduce \(\:{T}_{\text{m}\text{a}\text{x}}\) , and (iii) swaps of tasks among multiple vehicles to see if the objective can be improved. These procedures, integrated into the evolutionary process (creating a hybrid GA/local-search approach), contribute to faster convergence and higher quality solutions than using GA alone. Adaptation to Re-planning: In disaster response, conditions change continuously, so route plans need to be updated dynamically. The proposed system performs real-time re-planning when events such as new collapse reports, road blockage information, or vehicle failures (breakdown or withdrawal) occur. In practice, this means we take the vehicles’ current positions and any remaining tasks, then re-run the GA for a short duration (on the order of seconds to tens of seconds) to produce a re-optimised route plan reassigning tasks as needed. GAs are well-suited to successive re-optimisation because they can leverage parallel computation and the objective function is smooth/continuous with respect to small changes. In our approach, re-planning can be triggered either event-driven or at fixed time intervals, and the flexible reallocation of tasks allows the team as a whole to adapt with minimal increase in overall mission time. This ability to accommodate dynamic changes (a form of dynamic VRP) is a key feature of the cooperative GA approach. 3. Implementation and Data For proof of concept, we implemented the proposed framework in a simulation environment modelled on the Noto Peninsula region of Japan. The road network was derived from OpenStreetMap map data, which we extracted in GeoJSON format and converted to the node–edge graph as described in Section 2.2 . ( Fig. 3 shows a map of the Noto Peninsula region in Ishikawa Prefecture, Japan.) In the scenario, nine ground vehicles (three each from local government, fire, police, and SDF, totaling 9) patrol the affected area and collectively cover 70 inspection points. The starting bases (depots) for the 9 vehicles were dispersed near the area boundary so that the vehicles begin roughly spread out (for our simulation, depot coordinates were randomised; they could also be specified beforehand). Out of the 70 points, 20 points were hypothetically designated as locations where “building collapse has occurred,” and these were flagged as collapsed at the initial time. The collapse points were chosen randomly such that they were scattered without bias across the area. In the simulation logic, the actual determination of which points are collapsed would come from the RI/gNCDI analysis module. The proposed system was implemented in Python, utilising libraries such as NetworkX for graph processing and route finding. In the simulation, RI calculation and collapse detection were performed virtually (for implementation, we simply treated the 20 pre-selected points as collapsed); in future, this would be coupled with live video streaming input. The route optimisation by GA was performed with a custom GA engine; we tuned parameters such as population size, number of generations, crossover and mutation probabilities to ensure convergence. In our standard configuration, a population size of about 30–40 and a maximum of 60–80 generations were sufficient to reach a good solution. Although evaluating one individual involves computing the cost for a combination of 9 vehicles × 70 targets, this was done efficiently by referencing the precomputed distance-time table, making each generation evaluation very fast. As a baseline method, we implemented a greedy assignment heuristic with no cooperative optimisation. This Greedy method is a simple strategy wherein, at each step, it evaluates for each vehicle–target pair the increase in the objective function if that target were assigned next to that vehicle, and then selects the vehicle–target pair that yields the smallest cost increase. This assignment is repeated until all targets have been allocated, and finally each vehicle’s return leg to its depot is added to complete the routes. Because the objective function in our framework includes a weighting for collapsed points, the Greedy method does implicitly give some priority to collapsed sites (i.e. it considers not only distance but also the benefit of reaching a collapse sooner when choosing the next assignment). The Greedy method is computationally cheap and produces routes very quickly; however, due to its myopic, step-by-step nature, it often results in globally inefficient tours and uneven load distribution among vehicles (e.g. one vehicle may end up with a disproportionately long route). For evaluation, we compare the proposed method (Coop-GA) against the Greedy method in terms of key performance indicators such as the total time to complete coverage, the rapidity of collapse detection, and the balance of workload among vehicles. We also conduct comparisons on small-scale problems against brute-force optimal solutions, to verify how close the GA solutions are to the true optimum. Additionally, we simulate a scenario where some vehicles drop out during the mission and observe the effect of dynamic re-planning on route reconfiguration. As a visualisation, we generated a plot of all vehicle routes and survey points on the map (see Fig. 4) to illustrate the coverage and the area each vehicle is responsible for. Using the above setup, we perform a comprehensive evaluation of the framework’s operation. 4. Results 4.1 Performance Comparison: Greedy vs Cooperative GA First, we compare the route assignment results produced by the Greedy heuristic and the cooperative GA (Coop-GA) method. For the standard scenario (9 vehicles and ~ 70 target points, including 20 collapsed sites), Table 1 summarises the performance metrics of the plans obtained by both methods. Under the Greedy method, the route for the last vehicle became very long, and the overall completion time \(\:{T}_{\text{m}\text{a}\text{x}}\) was about 4.38 hours. In contrast, Coop-GA was able to reallocate tasks and shorten routes through genetic optimisation, reducing \(\:{T}_{\text{m}\text{a}\text{x}}\) to about 3.25 hours – a reduction of roughly 1.13 hours (25%). The total sum of arrival times at all collapsed sites was also improved: from approximately 18.99 vehicle-hours under Greedy (the sum of times at which each collapsed site was reached by the vehicles) down to 15.99 hours with Coop-GA, a ~ 16% reduction, indicating that collapsed building locations were incorporated earlier in the patrol order. Furthermore, the standard deviation of vehicle mission times (each vehicle’s total route duration) decreased from 1.48 hours with Greedy to 1.19 hours with Coop-GA, meaning the workload became more evenly distributed. In other words, under Greedy one particular vehicle ended up driving for a significantly longer time than the others, whereas Coop-GA produced a plan in which all vehicles finished their assignments in roughly equal times. Additionally, the overall objective function value \(\:J\) (as defined in Section 2.3 ) improved from 472.1 for Greedy to 352.7 for Coop-GA, about a 25% improvement (lower is better). These results confirm that the proposed cooperative GA method outperforms the greedy assignment in terms of shorter completion time, quicker discovery of collapsed buildings, and more balanced resource use. Table 1 Comparison of performance metrics for the standard scenario (9 vehicles, ~ 70 targets including 20 collapses). Method \(\:{T}_{\text{m}\text{a}\text{x}}\) (h) Sum of collapse arrival times (h) Std. dev. of vehicle times (h) Objective \(\:J\) Greedy 4.38 18.99 1.48 472.1 Coop-GA 3.25 15.99 1.19 352.7 In qualitative terms, the Greedy method’s distance-prioritised assignments caused a few faraway targets to all be left to one final vehicle, forcing that vehicle to make a large detour around almost the entire region. By contrast, Coop-GA’s genetic optimisation allowed task exchanges and route restructuring along the way, so far-off targets were split among multiple vehicles – resulting in a plan in which “no target was left behind” and everyone shared the work. Some collapsed sites were also deferred until later in the Greedy plan, whereas Coop-GA, due to the collapse-priority term in the objective, tended to schedule collapsed-site visits earlier in the route order. 4.2 Near-Optimality and Computation Time Next, we evaluated how close the Coop-GA solutions are to the true optimal solution on smaller problem instances, and examined the computation cost. We limited the problem to 3 vehicles and \(\:K=6,7,8\) targets, and performed an exhaustive search to find the optimal solution for each (since a brute-force search is only tractable at this small scale). We then compared the objective values of the Coop-GA solution and the Greedy solution against the optimum. The results showed that the Coop-GA’s objective value was within only a few percent of the exhaustive optimum, whereas the Greedy solution was substantially worse. For example, in the case of \(\:K=6\) targets, if we normalise the exhaustive optimum’s objective value to 100%, the Coop-GA solution was about ~ 102%, while the Greedy solution was around ~ 120%. For \(\:K=7\) and \(\:K=8\) , the Coop-GA remained within ≲105% (near-optimal), whereas the Greedy solutions were significantly higher (in one example between ~ 115–140% of the optimum). In particular, when comparing the completion time \(\:{T}_{\text{m}\text{a}\text{x}}\) , Greedy tends to be considerably larger than the optimal (because one vehicle’s route becomes excessively long), whereas Coop-GA, by suppressing the load of the slowest vehicle, achieved a \(\:{T}_{\text{m}\text{a}\text{x}}\) close to the optimal value. Also, regarding the speed of reaching the first collapse point, Coop-GA—thanks to the collapse-weighting in the objective—visits collapsed targets in an order close to the optimal strategy, whereas Greedy (which overly prioritises distance) often delays some collapses until later. From a computation cost perspective, the Greedy algorithm decides assignments one task at a time, so the number of evaluations grows roughly on the order of \(\:O\left(M\times\:K\right)\) (a quadratic function of number of vehicles \(\:M\) and targets \(\:K\) ). On the other hand, the Coop-GA evaluates POP individuals per generation and runs for GEN generations, so its complexity is approximately \(\:O\left(\text{POP}\times\:\text{GEN}\right)\) . In practice, POP and GEN can be scaled roughly linearly with problem size; in our setting, POP ≈ 30 and GEN ≈ 60 were sufficient, and as \(\:K\) increased, the runtime scaled approximately linearly. Empirically, we found that even for a problem of 9 vehicles and 70 targets, the route optimisation could converge in on the order of seconds to at most tens of seconds. By contrast, exhaustive search suffers combinatorial explosion: even at \(\:K=8\) it had to evaluate over 1.8 million possibilities, and at \(\:K=70\) the solution space would be astronomically large (on the order of 10^110 combinations), making brute force infeasible. Therefore, the framework is practically justified in that it can obtain near-optimal solutions very quickly. With sufficient computing resources, one could further accelerate this by parallelising or distributing the GA computations, allowing even larger problem sizes to be handled efficiently. Table 2 Exhaustive optimum vs. Coop-GA vs. Greedy for small problems (3 vehicles × \(\:K=6,7,8\) targets). (Exhaustive best is normalised to 100%.) Vehicles Targets \(\:K\) Exhaustive Best (100%) Coop-GA (% of optimum) Greedy (% of optimum) Exhaustive search evaluations 3 6 100% ≈ 102% ≈ 120% 20,160 3 7 100% ≲ 105% (near-optimal) > Coop (e.g. 115–140%) 181,440 3 8 100% ≲ 105% (near-optimal) > Coop (e.g. 115–140%) 1,814,400 4.3 Dynamic Re-planning Scenario In the latter part of the simulation, we tested a scenario where a vehicle becomes incapacitated partway through the mission (to simulate a breakdown or loss of communication), and assessed the effect of dynamic task reallocation by Coop-GA. Specifically, at the 50% mark of the mission, we removed one vehicle, and invoked the Coop-GA re-optimisation to redistribute that vehicle’s remaining targets to the others. Under the Greedy (non-cooperative) approach, the area that had been assigned to the dropped vehicle initially remained unvisited, and eventually the other vehicles had to cover those targets sequentially on their own – as a result, the overall completion time \(\:{T}_{\text{m}\text{a}\text{x}}\) was significantly extended. In contrast, with the Coop-GA approach, the dropout event was detected and GA was immediately re-run to reassign the remaining tasks of the lost vehicle to the others; as a result, the increase in \(\:{T}_{\text{m}\text{a}\text{x}}\) was kept very small. In other words, the cooperative system quickly reallocated the workload, so that even with one vehicle missing the team’s delay in finishing was minimised. The qualitative comparison in Table 1 also rated Coop-GA superior in terms of “scalability & fault tolerance,” noting that when a vehicle dropped out, the increase in \(\:{T}_{\text{m}\text{a}\text{x}}\) was much smaller for Coop-GA, reflecting its advantageous re-planning capability. Taken together, the results indicate that the proposed cooperative multi-vehicle system exhibits significantly higher performance than a static greedy heuristic. It enables rapid wide-area coverage and swift arrival at critical (collapsed) locations, allows efficient division of labour with no redundant overlap across multiple agencies, and also demonstrates better robustness to changing conditions. 4.4 Damage Detection Performance We conducted an experiment (simulating the Noto Peninsula damage scenario) to evaluate the collapse classification performance of the proposed gNCDI + GA image analysis method. As illustrated in Fig. 5 , we applied our model to features extracted from images labeled “damage present” and “no damage” respectively. The resulting confusion matrix of the classification was: TP = 9, FP = 2, TN = 35, FN = 1. In this test, the method achieved a Precision = 0.8182 and Recall = 0.9000, yielding an overall F1 score = 0.8571. In other words, on this limited validation set the method detected 90% of collapsed buildings without missing any, while keeping false alarms to about 18%. In quantitative terms, both precision and recall improved compared to a conventional method that uses a simple RI threshold, and a significant improvement in F1 was observed. For example, cases of collapsed houses covered by blue tarpaulins (low red component), or cases with only wall cracking and little rubble, were difficult for RI to detect, but the proposed method correctly identified the damage by leveraging colour-difference information. The sample detection results in Fig. 5 also show that our gNCDI + GA approach can pick up damage sites that the RI-based method missed. These findings indicate that the proposed method enables high-accuracy damage judgement even when only limited image data are available, and that qualitatively it expanded the range of damage patterns that can be detected, enhancing its practical utility in the field. Considering both qualitative and quantitative evaluation results, we gained confidence that this framework can automatically generate a reliable damage map within a few hours of an earthquake. In addition to confirming its rapidity and high spatial resolution advantage over traditional satellite or aerial imagery approaches, we also identified practical operational benefits of our method, such as inherent data privacy (using ground-level imagery rather than broad aerial surveillance) and a reduction in false detections caused by red-colored artifacts. These advantages, as well as remaining challenges, are discussed in the next section. 5. Discussion 5.1 Advantages and Limitations Advantages Balance of speed and accuracy: Whereas satellite-derived damage maps are often available only the day after an event or later, our framework can produce a detailed, ground-level damage map within a few hours. In simulation, 9 vehicles achieved full-area coverage in a little over 3 hours, which is less than half the time required by a single-vehicle operation (approximately inversely proportional to the number of vehicles). High collapse-detection performance: For EMS-98 D5 (collapse) detection, the gNCDI-based approach can be expected to achieve ≈ 90% recall, with false detections held to ≈ 20% or less (18% in our evaluation). Because the thresholding prioritizes recall to minimize misses, some false positives do occur; however, since missing a collapse is more critical, this trade-off is acceptable in practice. Operational benefits: A cloud-based architecture standardizes JSON sharing across agencies and maintains a single, real-time damage map. By integrating learning-free, lightweight image analysis (gNCDI) with GA-based vehicle route optimization, the framework enhances immediacy, scalability, and field deployability. Limitations (from the perspective of gNCDI-based image classification) Sensitivity to damage grade: As an RGB color-difference index, gNCDI is highly reliable for severe collapses where exposed timber/soil is prominent, but it tends to under-detect partial/moderate damage (EMS-98 D3/D4, e.g., wall cracking) where color changes are subtle. Imaging-condition dependence: Performance may degrade under nighttime/deep shadow, mixed lighting/auto white balance, motion blur, heavy dust/smoke, or color casts. While false positives from uniformly red structures (e.g., red roof tiles or painted walls) are reduced compared to simple RI, residual confounders remain when pigments, banners, rust, or background soils exhibit similar chromatic characteristics. Domain shift and threshold tuning: Differences in camera models, compression, and weather can alter color reproduction, causing the optimal threshold to drift. Per-scene calibration (using a simple reference image or a few samples) is recommended. Occlusion and visibility constraints: Vegetation or adjacent structures may occlude rubble from ground cameras. This can be mitigated by complementing ground surveys with UAV/helicopter observations and by multi-pass coverage from varied viewpoints. 5.2 Redness Index Sensitivity and Calibration The sensitivity of the Redness Index (and by extension gNCDI) can vary with local building materials and imaging conditions. For example, in regions with many wooden houses, a collapse results in conspicuous brownish-red timber and earth dust, yielding a strong RI signal; in contrast, in modern urban areas dominated by reinforced concrete, a collapse produces mostly gray debris and RI may respond weakly. Therefore, it is important to perform pre-calibration to set appropriate thresholds or colour corrections tailored to the region. Moreover, since imagery in bright midday conditions will have different colour characteristics than images taken in the evening or at night, applying brightness or colour temperature adjustments as needed can help maintain detection accuracy. In our experiments, we found that by using gNCDI (which generalises colour differences) instead of the simple redness index, we achieved better recall, precision, and F1 scores across varying conditions. This suggests that the generalised colour-difference approach is inherently more robust to differences in lighting and scene composition. 5.3 Load Balancing and Redundancy Control A major outcome of introducing the cooperative GA was the significant equalisation of task load among the multiple vehicles. Under Greedy assignment, it was common for one vehicle to end up with a concentration of targets at the end, leading to a large standard deviation in travel times [14]. This not only overburdens that particular vehicle, but also means the entire team’s response is prolonged while waiting for that single slowest vehicle to finish. Coop-GA, by including a deviation penalty in its objective, adjusted the assignment so that each vehicle’s distance/time is roughly even, yielding a plan where everyone can finish in step together. This is desirable also from the perspective of distributing human workload on the ground (more equitable fatigue levels). Regarding redundancy (overlapping coverage), by scenario design each target is only visited once, so the theoretical direct overlap is zero. However, inefficient routing can lead to indirect redundancy in the form of unnecessary detours. In the Greedy method, the order in which vehicles are assigned targets sometimes caused some vehicles to take long detours (effectively a kind of redundant travel). Coop-GA, through inter-vehicle task swaps and intra-route 2-opt optimisation, was able to compress this wasted travel. As a result, the proportion of redundant distance out of total distance driven was slightly reduced, indicating that cooperative GA is advantageous in curbing waste of transportation resources as well. 5.4 Behaviour under Vehicle Dropout and Re-planning As noted above, the Coop-GA approach possesses adaptability to dynamic changes such as vehicle dropout or new task arrival. A vehicle dropout (failure or loss of communication) is a very realistic possibility in the field, and traditional fixed plans cannot smoothly reallocate the remaining tasks. With Greedy assignment, the plan is decided once at the start and lacks flexibility, so if a dropout occurs, an additional delay is incurred while other vehicles eventually cover the orphaned tasks. In contrast, Coop-GA can immediately re-run its optimisation upon detecting a dropout, reassigning the lost vehicle’s tasks to the remaining vehicles such that the increase in \(\:{T}_{\text{m}\text{a}\text{x}}\) is minimised. This difference arises from the presence (or absence) of a real-time route re-optimisation capability, which is a major strength of our framework. In actual field operations, there are limitations in communication infrastructure and computing resources, but it is feasible to perform periodic re-computations on the order of every few seconds. Having multiple vehicles flexibly adjust their coverage areas in cooperation greatly enhances robustness under the uncertain and evolving conditions of a disaster. 5.5 Integration with Other Methods and Extensibility The proposed framework can function complementarily with existing damage assessment methods. For example, combining it with broad-area aerial imaging (e.g. from reconnaissance aircraft) or detailed assessments by structural experts would enable a more reliable and comprehensive understanding of damage. The damage maps automatically generated by our system can be integrated as a real-time layer into the common operational picture at disaster response headquarters, aiding decisions on where to deploy resources. Cross-referencing it with next-day satellite-based macro-scale damage maps could help to fill in any missed spots and to verify potential false positives. The framework itself is also highly generalisable and could be applied to other types of disasters such as heavy rain, typhoons, or floods. By swapping in different indices or sensor inputs appropriate to the disaster (for instance, a thermal infrared index for detecting burnt buildings in wildfires), one can utilise the ground-vehicle platform for rapid damage assessment across a wide range of scenarios. Thus, the approach offers broad applicability by adapting its indices and sensor payloads to different disaster types. 6. Conclusion Acknowledging the operational reality that, in the immediate aftermath of a disaster, there is no time to build or retrain deep learning models, this study presented a comprehensive framework that achieves wide-area, real-time damage assessment. The framework integrates collapsed-building detection from in-vehicle imagery using gNCDI, multi-agent route optimization via a Genetic Algorithm (GA), and inter-agency coordination through cloud-based JSON distribution. We implemented and validated the framework using both simulation and real-world footage. In a Noto Peninsula–style evaluation, coordination of nine vehicles indicated the potential to generate a wide-area damage map within a few hours. The gNCDI-based collapse detection achieved Precision = 0.8182, Recall = 0.9000, and F1 = 0.8571 (TP = 9, FP = 2, TN = 35, FN = 1), outperforming a simple RI-threshold method quantitatively. Compared with a greedy assignment, the GA produced more efficient patrol plans, shortening the overall completion time T max by about 25% while balancing workload, and converged to a near-optimal solution within seconds despite a search space on the order of 10 110 combinations. Moreover, even under dynamic changes such as vehicle dropout or road closures, rapid reallocation maintained full coverage; false detections remained at a level suppressible by subsequent human verification. Taken together, the framework is superior to conventional satellite/aerial approaches in both speed and spatial resolution and can effectively support early rescue and recovery. Looking ahead, we aim to evolve it into a standard, multi-hazard platform applicable to complex disasters (e.g., earthquakes and typhoons) that contributes to damage reduction and the protection of human life. The authors sincerely hope this work will help accelerate life-saving efforts. References Gómez-Chova, Luis, et al. “Multimodal classification of remote sensing images: A review and future directions.” Proceedings of the IEEE 103.9 (2015): 1560–1584. Tu, Jihui, et al. “Automatic building damage detection method using high-resolution remote sensing images and 3D GIS model.” ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences 3 (2016): 43–50. Tilon, Sofia, et al. “Post-disaster building damage detection from earth observation imagery using unsupervised and transferable anomaly detecting generative adversarial networks.” Remote Sensing 12.24 (2020): 4193. Monfort, Daniel, Caterina Negulescu, and Myriam Belvaux. “Remote sensing vs. field survey data in a post-earthquake context: Potentialities and limits of damaged building assessment datasets.” Remote Sensing Applications: Society and Environment 14 (2019): 46–59. Suppasri, Anawat, et al. “The 2024 Noto Peninsula earthquake: Preliminary observations and lessons to be learned.” International Journal of Disaster Risk Reduction 110 (2024): 104611. 2024 Noto earthquake bosaiXview. https://xview.bosai.go.jp/view/index.html?appid=41a77b3dcf3846029206b86107877780 (accessed 16 Jan 2025). Mizui, Y., Fujiwara, H. Estimate the amount of disaster waste disposal work using in-vehicle camera images – a case study in Hitoyoshi City, Kumamoto Prefecture. Journal of Disaster Research , 16(7), 2021, pp. 1061–1073. https://doi.org/10.20965/jdr.2021.p1061. Van der Meer, Freek. “Remote-sensing image analysis and geostatistics.” International Journal of Remote Sensing 33.18 (2012): 5644–5676. Ma, Lei, et al. “Deep learning in remote sensing applications: A meta-analysis and review.” ISPRS Journal of Photogrammetry and Remote Sensing 152 (2019): 166–177. Escadafal, R., Belghit, A., & Ben-Moussa, A. Indices spectraux pour la télédétection de la dégradation des milieux naturels en Tunisie aride, in : Guyot G. (ed.), Mesures physiques et signatures en télédétection: Sixième symposium international; actes; 17–21 Janvier 1994, Val d’Isère, France: Proceedings , CNES (Centre National d’Etudes Spatiales), Paris, 1994, pp. 253–259. Escadafal, R., & Huete, A. Étude des propriétés spectrales des sols arides appliquée à l’amélioration des indices de végétation obtenus par télédétection [Improvement in remote sensing of low vegetation cover in arid regions by correcting vegetation indices for soil “noise”]. Comptes Rendus de l’Académie des Sciences, Série II 312(11), 1991, pp. 1385–1391. Shiraishi, H., & Usuda, Y. (2025). Real-Time Building-Damage-Extraction Technology from Ground-Based Video Footage Using Normalized Difference Red/Green Redness Index . Geomatics and Environmental Engineering, 19(1), 143–159. https://doi.org/10.7494/geom.2025.19.1.143 Taniguchi, E., Yamada, T., & Kakimoto, Y. Probabilistic vehicle routing and scheduling with variable travel times. IFAC Proceedings , 33(9), 2000, pp. 33–38. https://doi.org/10.1016/S1474-6670(17)38119-3. Kefi, M., & Ghedira, K. “A multi-agent model for the Vehicle Routing Problem with Time Windows.” WIT Transactions on The Built Environment 75 (2004). Zidi, Kamel, et al. “Distributed genetic algorithm for disaster relief planning.” International Journal of Computers Communications & Control 8.5 (2013): 769–783. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Published Journal Publication published 18 Nov, 2025 Read the published version in Natural Hazards Research → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7390282","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":501398431,"identity":"965eba88-caf8-4a3f-8aa0-fa928029ad42","order_by":0,"name":"Haruhiro Shiraishi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA80lEQVRIiWNgGAWjYPCCA0DMAyJtgAzGxgOkaEkDaWkgXgsQHIZzcQLzGclHN/xguCNvPiP34OGKivN2a9sPA22psYnGpUXmRlrazR6GZ4ZzbuQlHDxz5nbytjOJQC3H0nIbcGiRkMgxu8HDcJhxhkSOwcHGttvJZgeAWhgbDuPVcvMPw2F7qJZzyWbnHxLWchtoSyJUywE7sxuEbOF5lnZbxuBw8gyeNwYHG84kJ5jdANqSgM8v7MnHbr6pOGw7gz3H+GNDhZ292fn0hw8+1Njg1MIgkAAkDBD8RLDKBFzKQYD/ACrfHp/iUTAKRsEoGJkAACzTaPbl2HSaAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-2585-9335","institution":"National Research Institute for Earth Science and Disaster Prevention, University of Tsukuba","correspondingAuthor":true,"prefix":"","firstName":"Haruhiro","middleName":"","lastName":"Shiraishi","suffix":""},{"id":501399063,"identity":"6f72e66c-cee1-4163-a7a9-49d2749c3b90","order_by":1,"name":"Yuichiro Usuda","email":"","orcid":"https://orcid.org/0000-0002-8660-7270","institution":"National Research Institute for Earth Science and Disaster Prevention, University of Tsukuba","correspondingAuthor":false,"prefix":"","firstName":"Yuichiro","middleName":"","lastName":"Usuda","suffix":""}],"badges":[],"createdAt":"2025-08-17 04:05:09","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7390282/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7390282/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1016/j.nhres.2025.11.003","type":"published","date":"2025-11-19T00:00:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":89456022,"identity":"1568e6e5-cf9c-471f-b9bd-368daac8a3ad","added_by":"auto","created_at":"2025-08-20 06:56:19","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":631602,"visible":true,"origin":"","legend":"\u003cp\u003eConceptual illustration of “drive-by” ground imaging vs. traditional on-site survey.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/492c0f49d1e294e57f019c6d.png"},{"id":89454987,"identity":"ddc68d6f-7cb0-497f-b081-b9a17746c6e7","added_by":"auto","created_at":"2025-08-20 06:48:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":90150,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture of the proposed framework.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/cfae8c8043e6d68c6082c1fe.png"},{"id":89454983,"identity":"8f81991e-641f-4fae-a32a-25a83315fcb1","added_by":"auto","created_at":"2025-08-20 06:48:19","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":123851,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eMap of the Noto Peninsula region (Ishikawa, Japan), showing the road network and simulated survey points.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/b34ff1766fef983e258f1b4e.png"},{"id":89456349,"identity":"cce51e70-0422-4074-8bc5-406fd5aa4266","added_by":"auto","created_at":"2025-08-20 07:04:19","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":486406,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eRoute planning results comparison: (a) Cooperative GA–optimised routes, (b) Greedy heuristic routes, and (c) legend indicating vehicle paths and target points.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4a1.png","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/af93852ef79c926c84bdb6e6.png"},{"id":89456351,"identity":"f668938a-05ac-48a7-ad45-64ae95c63e38","added_by":"auto","created_at":"2025-08-20 07:04:19","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":457116,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eExample of damage detection results. The proposed gNCDI+GA method (right) correctly identifies collapsed building damage in cases that the simpler RI-based method (left) missed.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/47baf8a9831a306c905b1236.png"},{"id":96411218,"identity":"28f5c269-7e63-42bd-8def-4fa2d2af9e35","added_by":"auto","created_at":"2025-11-20 18:45:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2449208,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7390282/v1/5aca1207-d196-4083-8ac3-be52e2d0cdc1.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eA Practical Framework for Rapid Earthquake Damage Estimation through Multi-Vehicle Coordination and Inter-Agency Collaboration: Integrating Genetic Algorithms with RGB-Based Image Analysis\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAfter a strong earthquake, emergency responders need to obtain a rapid and accurate overview of building damage. Information on the locations of collapsed buildings and the severity of damage is crucial for guiding search-and-rescue efforts and efficient resource distribution. Traditionally, remote sensing methods using satellite imagery or aerial photographs have been employed for damage assessment [1]. However, these approaches suffer from temporal delays and practical constraints. For example, satellites have limited imaging opportunities and are easily hindered by cloud cover, and although helicopters or drones can capture high-resolution data, they typically cannot cover broad areas quickly and lack immediacy. Advanced image analysis techniques (such as detailed 3D map reconstruction [2] or complex anomaly detection algorithms [3]) are computationally intensive and often unsuitable for immediate post-disaster response. As a result, in practice ground visual surveys by municipal engineers, fire services, and police are a primary information source[4]\u0026mdash;but these are labour-intensive and tend to be siloed by agency, leading to fragmented information. To bridge this gap, we propose a framework that combines multiple ground vehicles with a centralized system to automatically generate a reliable damage map within a few hours of an earthquake. This idea draws inspiration from recent technological trends: during the 2024 Noto Peninsula earthquake[5], for instance, vehicles equipped with 360\u0026deg; street-level cameras[6] were deployed to drive around the affected area and collect panoramic imagery. Such \u0026ldquo;drive-by\u0026rdquo; imaging can acquire close-range, high-resolution photos far more rapidly than foot surveys (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Moreover, coordinating vehicles from multiple agencies in a collaborative manner could potentially cover a wide area in a short time. Indeed, prior research also suggests the usefulness of ground mobile imaging. For example, Mizui and Fujiwara (2021) analysed vehicle dashboard camera footage to estimate disaster waste volume, demonstrating that a quick survey using in-vehicle video after the Hitoyoshi City floods enabled an accurate assessment of debris to be removed [7]. This indicates that rich information on damage extent can be extracted from footage captured by moving vehicles, which motivates our use of ground-level image data in this work.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAdvances in automated image analysis technology further support this research. Traditional computer-vision methods [8] (e.g. colour indices, texture analysis, edge detection) have been used to classify building damage, but in recent years deep learning approaches like Convolutional Neural Networks (CNNs) have become predominant [9]. However, deep learning requires large amounts of training data and computational resources, and the models act as \u0026ldquo;black boxes\u0026rdquo; with opaque decision logic. In the immediate aftermath of a disaster, a method that, even if simple, is rapid, explainable, and easily adjustable in the field is preferable. We focus on a simple colour-based index known as the Redness Index (RI)[10][11][12]. RI is a normalized difference computed from red and green colour intensities, defined by the following formula (1):\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\text{R}\\text{I}=\\frac{R-G}{\\hspace{0.17em}R+G\\hspace{0.17em}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eOriginally proposed as an index to correct vegetation measurements in arid regions[10][11], RI is particularly responsive to materials with strong red reflectance and weak green reflectance. We hypothesise that this spectral characteristic can serve as an indicator of severe structural damage: when a building collapses, structural timber becomes exposed and dust from roof tiles or earth plaster tends to increase the relative red reflectance. Because it requires no extensive training data and can be implemented as simple per-pixel calculations with thresholding, RI is well-suited to real-time processing in the field. Shiraishi and Usuda (2025) proposed the concept of using RI (a Normalized Red\u0026ndash;Green Index) for building damage extraction and showed that RI could detect nearly all completely-collapsed buildings (damage grade D5 on the European Macroseismic Scale EMS-98) in the Noto Peninsula earthquake data[12]. In their preliminary validation, 14 out of 15 D5-level collapsed buildings were correctly detected by an RI threshold, and 14 out of 16 non-collapsed buildings were correctly identified as undamaged (most false alarms were intact buildings with red-colored roofs or walls)[12]. Building on the high recall and accuracy demonstrated by this simple colour index, our framework generalises RI into a generalized Normalized Chromatic Difference Index (gNCDI) (see Eq.\u0026nbsp;(\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) below) and places this at the core of a large-scale damage mapping approach.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{g}\\text{e}\\text{n}\\text{e}\\text{r}\\text{a}\\text{l}\\text{i}\\text{z}\\text{e}\\text{d}\\:\\text{N}\\text{o}\\text{r}\\text{m}\\text{a}\\text{l}\\text{i}\\text{z}\\text{e}\\text{d}\\:\\text{C}\\text{h}\\text{r}\\text{o}\\text{m}\\text{a}\\text{t}\\text{i}\\text{c}\\:\\text{D}\\text{i}\\text{f}\\text{f}\\text{e}\\text{r}\\text{e}\\text{n}\\text{c}\\text{e}\\:\\text{I}\\text{n}\\text{d}\\text{e}\\text{x}\\:\\left(\\text{g}\\text{N}\\text{C}\\text{D}\\text{I}\\right)=$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:a\\cdot\\:\\frac{R-G}{\\hspace{0.17em}R+G\\hspace{0.17em}}\\hspace{0.33em}+\\hspace{0.33em}b\\cdot\\:\\frac{G-B}{\\hspace{0.17em}G+B\\hspace{0.17em}}\\hspace{0.33em}+\\hspace{0.33em}c\\cdot\\:\\frac{B-R}{\\hspace{0.17em}B+R\\hspace{0.17em}}+d・\\frac{R}{\\hspace{0.17em}R+G+B\\hspace{0.17em}}+e・\\frac{G}{\\hspace{0.17em}R+G+B\\hspace{0.17em}}+f・\\frac{B}{\\hspace{0.17em}R+G+B\\hspace{0.17em}}+\\:g\\text{・}\\text{S}+h\\text{・}\\text{c}\\text{o}\\text{s}\\left(2{\\pi\\:}\\text{h}\\right)+\\text{i}\\text{・}\\text{s}\\text{i}\\text{n}\\left(2{\\pi\\:}\\text{h}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eHere, RGB equals R, G, B in a linear intensity space. Let HSV be (H, S, V) with H in [0,1) and S in [0,1].\u003c/p\u003e\u003cp\u003eAnother challenge is the coordinated route planning for multiple vehicles. In a large-scale disaster, various organisations (municipal authorities, fire brigades, police, Self-Defense Forces, etc.) each patrol their assigned areas to assess damage. Without sufficient coordination, however, survey efforts may overlap or miss certain areas. To enable efficient collaboration, we model the problem as a time-constrained multi-agent tour on a road network, and apply a Genetic Algorithm (GA) for optimising the routes of multiple vehicles. Vehicle routing problems (VRPs) and their variants have been widely studied in the logistics domain, and their application to disaster response has also been explored. For example, Taniguchi et al. (2000) investigated a probabilistic, time-varying travel time model accounting for uncertainty in disaster situations [13]. Kefi and Ghedira (2004) proposed a cooperative multi-agent solution for the VRP with time windows[14]. Zidi et al. (2013) applied a GA to disaster relief vehicle routing, dynamically incorporating new requests (rescue calls, etc.) and interruption events (e.g. road closures) to achieve a solution that maximised the number of affected people reached while minimising total travel cost [15]. Our approach follows the same principle, performing a multi-objective optimisation that maximises survey coverage (number of damaged buildings detected) while minimising the overall completion time for all vehicles.\u003c/p\u003e\u003cp\u003eWith regard to multi-agency information sharing, a unified command system such as the Incident Command System (ICS) is internationally adopted. Our proposed centralized coordination server aggregates and analyzes data from all vehicles regardless of their parent organisation, and dispatches task assignments or route changes to each vehicle in real time. This centralized control enables all agencies to operate with a common, up-to-date damage map and progress status, while maintaining agility in the field.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn summary, the framework presented in this research is a novel rapid damage assessment system that combines ground-level image analysis with cooperative vehicle routing. The main contributions of this work are as follows:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eMethodology: We propose an earthquake damage mapping approach that integrates a simple colour-based damage detection technique with a multi-vehicle route optimisation strategy.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eSystem Design: We design a cloud-based, real-time information-sharing system for inter-agency collaboration, using a standardised JSON data format for geospatial damage reporting.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eValidation: We verify the effectiveness of the proposed approach through simulation of a scenario in Noto Peninsula, and conduct a performance comparison with conventional methods (e.g. a greedy heuristic).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003ePractical Considerations: We discuss issues for real-world deployment, including ethics, privacy, and data reliability.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eThe remainder of this paper is organized as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e describes the proposed framework in detail, including the image-based damage detection and cooperative vehicle routing methods. Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e3\u003c/span\u003e outlines the implementation and dataset used in our simulation. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the evaluation results, and Section \u003cspan refid=\"Sec13\" class=\"InternalRef\"\u003e5\u003c/span\u003e provides discussion of advantages, limitations and practical implications. Finally, Section \u003cspan refid=\"Sec19\" class=\"InternalRef\"\u003e6\u003c/span\u003e offers concluding remarks.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eIn this section, we explain the core techniques of the proposed framework. First, we describe the image-based damage detection method using the gNCDI index. Next, we explain the road network data used for route planning and how hazardous areas are handled. Finally, we detail the design of the cooperative GA (Genetic Algorithm) \u0026mdash; including the chromosome representation, evolutionary operators, and the route optimisation objective function.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Damage Detection Using gNCDI and GA\u003c/h2\u003e\u003cp\u003eTo enable real-time damage interpretation on frames transmitted from vehicles in disaster areas, we combine the generalized Normalized Chromatic Difference Index (gNCDI) with a Genetic Algorithm (GA). gNCDI generalizes the conventional Redness Index (RI). While RI essentially uses only a single \u0026ldquo;red\u0026ndash;green difference,\u0026rdquo; our approach employs a 9-dimensional color feature vector. This increases sensitivity not only to reddish debris but also to cues that RI tends to miss, such as blue tarpaulins and fine cracks.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e9-Dimensional Features and Definition of gNCDI\u003c/h3\u003e\n\u003cp\u003eLet RGB be R, G, B in a linear intensity space. Let HSV be (H, S, V) with H in [0,1) and S in [0,1]. Let eps\u0026thinsp;=\u0026thinsp;1e-6 be a small constant to avoid division by zero. Define the 9-dimensional feature vector x as:\u003c/p\u003e\u003cp\u003ex1 = (R - G) / (R\u0026thinsp;+\u0026thinsp;G\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex2 = (G - B) / (G\u0026thinsp;+\u0026thinsp;B\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex3 = (B - R) / (B\u0026thinsp;+\u0026thinsp;R\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex4\u0026thinsp;=\u0026thinsp;R / (R\u0026thinsp;+\u0026thinsp;G\u0026thinsp;+\u0026thinsp;B\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex5\u0026thinsp;=\u0026thinsp;G / (R\u0026thinsp;+\u0026thinsp;G\u0026thinsp;+\u0026thinsp;B\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex6\u0026thinsp;=\u0026thinsp;B / (R\u0026thinsp;+\u0026thinsp;G\u0026thinsp;+\u0026thinsp;B\u0026thinsp;+\u0026thinsp;eps)\u003c/p\u003e\u003cp\u003ex7\u0026thinsp;=\u0026thinsp;S\u003c/p\u003e\u003cp\u003ex8\u0026thinsp;=\u0026thinsp;cos(2\u003cem\u003epi\u003c/em\u003eH)\u003c/p\u003e\u003cp\u003ex9\u0026thinsp;=\u0026thinsp;sin(2\u003cem\u003epi\u003c/em\u003eH)\u003c/p\u003e\u003cp\u003eWith coefficient vector w = [a, b, c, d, e, f, g, h, i] and bias b0, the gNCDI score is:\u003c/p\u003e\u003cp\u003egNCDI(x)\u0026thinsp;=\u0026thinsp;a\u003cem\u003ex1\u0026thinsp;+\u0026thinsp;b\u003c/em\u003ex2\u0026thinsp;+\u0026thinsp;c\u003cem\u003ex3\u0026thinsp;+\u0026thinsp;d\u003c/em\u003ex4\u0026thinsp;+\u0026thinsp;e\u003cem\u003ex5\u0026thinsp;+\u0026thinsp;f\u003c/em\u003ex6\u0026thinsp;+\u0026thinsp;g\u003cem\u003ex7\u0026thinsp;+\u0026thinsp;h\u003c/em\u003ex8\u0026thinsp;+\u0026thinsp;i*x9\u0026thinsp;+\u0026thinsp;b0\u003c/p\u003e\u003cp\u003eThe first three terms (opponent color differences) are robust to illumination changes and capture red\u0026ndash;green, green\u0026ndash;blue, and blue\u0026ndash;red contrasts. The next three terms (RGB proportions) reflect color composition relative to overall brightness. The last three terms embed hue on the unit circle via cos and sin and include saturation. As a result, the model responds not only to \u0026ldquo;redness\u0026rdquo; (RI-like debris) but also to blue\u0026ndash;red contrasts (e.g., blue tarps) and subtle chromatic irregularities associated with cracks.\u003c/p\u003e\u003cp\u003eLearning with GA (F1 Maximization)\u003c/p\u003e\u003cp\u003eWe use gNCDI(x) as a damage score and classify a pixel (or image) as \u0026ldquo;damaged\u0026rdquo; if it exceeds a threshold tau. During training, the coefficients w and bias b0 are jointly optimized by GA, and for each candidate solution we internally search the threshold that maximizes the F1-score based on the score distribution.\u003c/p\u003e\u003cp\u003eProcedure: (i) initialize a random population; (ii) evaluate each individual by computing precision and recall, then its F1 as fitness; (iii) keep elites via elitist selection; (iv) generate offspring by crossover (real-valued recombination of weights and bias); (v) apply mutation (small perturbations) to maintain diversity. This yields (w, b0, tau_star) tailored to local conditions even with a small calibration set.\u003c/p\u003e\u003cp\u003eDeployment (Real-Time Inference)\u003c/p\u003e\u003cp\u003eAt inference, compute x for all pixels in each frame, evaluate gNCDI(x) with the optimized w and b0, and extract pixels with scores\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;tau_star as damage-indicative pixels. If high-scoring pixels form clusters exceeding a set area, flag the location as a damage candidate; voting across consecutive frames further reduces false alarms. The method remains lightweight and easy to implement like RI-based approaches, while the 9-D formulation captures richer damage cues, improving detection for cases difficult for RI alone (blue tarps, fine cracks).\u003c/p\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Road Network Data and Hazardous Area Handling\u003c/h2\u003e\u003cp\u003eEfficient touring by multiple vehicles critically depends on leveraging road infrastructure data. In this study, the road network of the target region is provided in GeoJSON format, which we convert into a graph data structure for route planning. Specifically, each road segment (LineString in GeoJSON) is parsed into nodes and edges: every intersection or endpoint becomes a node, and connecting road segments are represented as edges. The length of each edge is calculated from the haversine distance of the corresponding road segment (the great-circle distance based on latitude/longitude), as an approximation of driving distance. We normalise node coordinates by rounding latitude/longitude to a consistent precision and merge any duplicate nodes as a preprocessing step to ensure a clean, connected graph.\u003c/p\u003e\u003cp\u003eSince road blockages are likely to occur immediately after an earthquake, we incorporate known hazardous areas into the route planning to exclude those segments. The framework allows optional input of hazard polygons (areas of concern such as potential sinkholes, structural collapses, or severely flooded zones). For each provided hazard polygon, the system checks each edge of the road graph to determine if it geometrically intersects or lies within the polygon. Any such edge is marked as impassable and removed from the graph. This ensures that the routing algorithm will automatically avoid paths through dangerous areas. After removing edges, any node that becomes isolated (no longer reachable from any other node) is removed from the set of targets to maintain overall graph connectivity.\u003c/p\u003e\u003cp\u003eWe fix the vehicle travel speed at ~\u0026thinsp;40 km/h (typical urban driving speed) for the simulation, converting graph distances to travel times by this speed. The primary route cost calculations use this travel time as the weight. In real operations, one should account for traffic obstructions or congestion, but for simplicity in our simulation we assume a constant speed (it is possible to pre-assign per-road penalties based on expected travel difficulty, if needed). We assume one depot (base) location for each agency, i.e. each of the 9 vehicles starts and ends at its designated depot. Depot coordinates can be provided as a set of GeoJSON Points; if no depots are specified, we randomly placed nine depots near the area boundary in our simulation.\u003c/p\u003e\u003cp\u003eThe points of interest (targets) represent the locations of buildings or areas to be inspected. We set approximately 70 target points uniformly sampled from the road network nodes, which can be thought of as representing the buildings to survey in the region. Furthermore, we marked about 20 of these as points where \u0026ldquo;building collapse occurred\u0026rdquo; in the scenario. The collapse points were selected at random (uniformly distributed across the area) in this study, but in practice they would be determined dynamically by the gNCDI analysis module in real time. In other words, the gNCDI processing module would flag locations with \u0026ldquo;possible\u0026rdquo; values as \u0026ldquo;suspected collapse,\u0026rdquo; which would then be shared with the route planning module. In our simulation, we simply fixed 20 points as collapsed at the start, but the framework is designed to operate under the assumption that the set of collapse flags is updated dynamically during the mission.\u003c/p\u003e\u003cp\u003eOnce the road graph and the depots/targets are prepared, we pre-compute a distance table to be used for route planning. We calculate the shortest-path distance on the graph between every pair of POIs (Points of Interest: roughly 70 targets\u0026thinsp;+\u0026thinsp;9 depots, ~\u0026thinsp;79 points in total) using Dijkstra\u0026rsquo;s algorithm, and convert those distances to travel times to form a time cost matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:T\\)\u003c/span\u003e\u003c/span\u003e. Any pair that is unreachable (for example, if two points are on separate islands with no connecting roads) is assigned an infinite distance. This preprocessing speeds up the cost evaluations and the optimisation procedure described next.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Cooperative GA: Chromosome Design and Objective Function\u003c/h2\u003e\u003cp\u003eFor route optimisation, we adopt a cooperative Genetic Algorithm (Cooperative GA, abbreviated Coop-GA). In this approach, a single individual (solution) represents the route assignments for all vehicles, and evolutionary computation is used to search for improved solutions. Below, we describe the design of the GA method used in this study.\u003c/p\u003e\u003cp\u003eChromosome Representation: One individual encodes the set of tour routes for all vehicles. Specifically, it is represented as a tuple concatenating the ordered list of assigned targets for each of the 9 vehicles (e.g. an individual could be encoded as a tuple like ((\u0026hellip;),(\u0026hellip;), \u0026hellip; ,(\u0026hellip;)), where each component list corresponds to one vehicle\u0026rsquo;s target visitation sequence). Each route is a sequence of departing from the vehicle\u0026rsquo;s depot, visiting its assigned targets in order, and returning to the depot. However, the depot is not explicitly included in the chromosome; each vehicle\u0026rsquo;s gene sequence consists only of target IDs. During evaluation, the depot of each vehicle is automatically prepended and appended to that vehicle\u0026rsquo;s target list to form the complete circuit.\u003c/p\u003e\u003cp\u003eInitial Population: Rather than purely random initialisation, the initial population of solutions is seeded with a heuristic solution to ensure both diversity and quality. First, we generate one heuristic route plan by assigning each target to the vehicle with the nearest depot, and ordering each vehicle\u0026rsquo;s targets roughly by nearest-neighbour (shortest incremental distance). This provides a plausible initial solution (we also generate the Greedy solution described later, and include it as well). We inject this heuristic solution (and the Greedy solution) into the initial population, while generating the remaining individuals with random task allocations and orders. This way, the initial population contains a few relatively good solutions while still maintaining broad variety.\u003c/p\u003e\u003cp\u003eFitness Evaluation (Objective Function): The GA\u0026rsquo;s evaluation function (fitness) is a composite cost metric for the set of routes across all vehicles. We define a weighted sum objective function that integrates three aims: (i) minimising the overall completion time, (ii) arriving at collapsed sites as early as possible, and (iii) balancing the workload among vehicles. Formally:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:J={w}_{1}\\hspace{0.17em}{T}_{\\text{m}\\text{a}\\text{x}}\\hspace{0.33em}+\\hspace{0.33em}{w}_{2}\\sum\\:_{i\\in\\:Collapsed}{t}_{i}\\hspace{0.33em}+\\hspace{0.33em}{w}_{3}\\hspace{0.17em}\\sigma\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:J\\)\u003c/span\u003e\u003c/span\u003e is the objective value for a given complete plan of routes for all vehicles. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e is the maximum time taken by any vehicle to complete its route (i.e. the time when the last vehicle finishes, also known as the makespan). \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{t}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the arrival time at collapse-flagged target \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e (and the summation is the total of all such arrival times across identified collapse points). \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\sigma\\:\\)\u003c/span\u003e\u003c/span\u003e is the standard deviation of individual vehicle mission times (the time from each vehicle\u0026rsquo;s departure to its return to depot). The weight coefficients \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{w}_{1},{w}_{2},{w}_{3}\\)\u003c/span\u003e\u003c/span\u003e are set according to the relative importance of these objectives. In our implementation, we assign weights such that reducing \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e is by far the top priority (i.e. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{w}_{1}\\)\u003c/span\u003e\u003c/span\u003e is orders of magnitude larger than the others), while also penalising delays in reaching collapsed sites and large disparities in vehicle workloads. In our standard setting, for example, we heavily weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e to ensure the fleet\u0026rsquo;s total completion time is minimised, and choose smaller but non-zero weights for the collapse arrival term and the time deviation term. In effect, this single scalar objective encodes a multi-objective optimisation that seeks to shorten the overall mission time (minimise \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e ), ensure prompt visits to collapsed-building locations (minimise the sum of collapse arrival times), and equalise the workload across vehicles (minimise the time standard deviation). As a result, the GA will assign higher fitness to solutions where no single vehicle is extremely delayed (due to the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e term), collapsed sites are visited early in the routes (due to the second term), and each vehicle finishes in roughly the same time (due to the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\sigma\\:\\)\u003c/span\u003e\u003c/span\u003e term).\u003c/p\u003e\u003cp\u003eSelection, Crossover, and Mutation: The GA\u0026rsquo;s core operations are similar to their application in other combinatorial optimisations. In each generation, selection is applied based on objective value (using techniques like roulette-wheel or tournament selection) so that better individuals are more likely to be chosen for reproduction. For crossover, we adopt a \u003cem\u003evehicle-wise route segment crossover\u003c/em\u003e. That is, for each of the 9 vehicles, we randomly extract a portion of the route (a contiguous subsequence of targets) from parent A and parent B, and combine them to form each vehicle\u0026rsquo;s route in the child. In this way, a single crossover operation recombines the routes of all vehicles simultaneously. After crossover, the child solution may contain duplicate visits or unassigned targets; we apply a repair operator to remove any duplicate target assignments and reinsert any missing targets, thereby maintaining a valid complete solution.\u003c/p\u003e\u003cp\u003eFor mutation, we introduce small random modifications to an individual to promote diversity. We combine multiple mutation operators in our implementation【15】, including:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eInter-vehicle relocation: Remove one task from the most heavily-loaded vehicle and insert it into the route of the least loaded vehicle. This adjusts the balance of task counts between vehicles, alleviating extreme disparities (directly reducing the standard deviation term in the objective).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eInter-vehicle swap: Randomly pick two vehicles and swap one task between their routes. By transferring a relatively out-of-the-way task to another vehicle, this operator explores the possibility of reducing overall distance/time.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIntra-route 2-opt shuffle: For each vehicle\u0026rsquo;s route, either swap the order of two visits that are very close to each other, or take a randomly selected segment of the route and reverse its order. This is analogous to the classic 2-opt method in the Traveling Salesman Problem, and serves as a local optimisation to shorten an individual vehicle\u0026rsquo;s route.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eAfter any mutation, a repair step is again performed to fix any inconsistencies (e.g. duplicate or missing tasks). By applying these operators (potentially multiple times per generation) in combination, the algorithm explores a wide search space while still guiding the population towards convergence.\u003c/p\u003e\u003cp\u003eLocal Search Enhancement: Once the GA solutions have converged to a reasonably good region, we further apply a local search to fine-tune and improve the solution. Specifically, we sequentially apply: (i) a 2-opt optimisation on each vehicle\u0026rsquo;s route to remove any unnecessary detours, (ii) a relocation of one task from the slowest vehicle to the fastest vehicle to directly reduce \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, and (iii) swaps of tasks among multiple vehicles to see if the objective can be improved. These procedures, integrated into the evolutionary process (creating a hybrid GA/local-search approach), contribute to faster convergence and higher quality solutions than using GA alone.\u003c/p\u003e\u003cp\u003eAdaptation to Re-planning: In disaster response, conditions change continuously, so route plans need to be updated dynamically. The proposed system performs real-time re-planning when events such as new collapse reports, road blockage information, or vehicle failures (breakdown or withdrawal) occur. In practice, this means we take the vehicles\u0026rsquo; current positions and any remaining tasks, then re-run the GA for a short duration (on the order of seconds to tens of seconds) to produce a re-optimised route plan reassigning tasks as needed. GAs are well-suited to successive re-optimisation because they can leverage parallel computation and the objective function is smooth/continuous with respect to small changes. In our approach, re-planning can be triggered either event-driven or at fixed time intervals, and the flexible reallocation of tasks allows the team as a whole to adapt with minimal increase in overall mission time. This ability to accommodate dynamic changes (a form of dynamic VRP) is a key feature of the cooperative GA approach.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Implementation and Data","content":"\u003cp\u003eFor proof of concept, we implemented the proposed framework in a simulation environment modelled on the Noto Peninsula region of Japan. The road network was derived from OpenStreetMap map data, which we extracted in GeoJSON format and converted to the node\u0026ndash;edge graph as described in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e. \u003cem\u003e(\u003c/em\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e \u003cem\u003eshows a map of the Noto Peninsula region in Ishikawa Prefecture, Japan.)\u003c/em\u003e In the scenario, nine ground vehicles (three each from local government, fire, police, and SDF, totaling 9) patrol the affected area and collectively cover 70 inspection points. The starting bases (depots) for the 9 vehicles were dispersed near the area boundary so that the vehicles begin roughly spread out (for our simulation, depot coordinates were randomised; they could also be specified beforehand). Out of the 70 points, 20 points were hypothetically designated as locations where \u0026ldquo;building collapse has occurred,\u0026rdquo; and these were flagged as collapsed at the initial time. The collapse points were chosen randomly such that they were scattered without bias across the area. In the simulation logic, the actual determination of which points are collapsed would come from the RI/gNCDI analysis module.\u003c/p\u003e\u003cp\u003eThe proposed system was implemented in Python, utilising libraries such as NetworkX for graph processing and route finding. In the simulation, RI calculation and collapse detection were performed virtually (for implementation, we simply treated the 20 pre-selected points as collapsed); in future, this would be coupled with live video streaming input. The route optimisation by GA was performed with a custom GA engine; we tuned parameters such as population size, number of generations, crossover and mutation probabilities to ensure convergence. In our standard configuration, a population size of about 30\u0026ndash;40 and a maximum of 60\u0026ndash;80 generations were sufficient to reach a good solution. Although evaluating one individual involves computing the cost for a combination of 9 vehicles \u0026times; 70 targets, this was done efficiently by referencing the precomputed distance-time table, making each generation evaluation very fast.\u003c/p\u003e\u003cp\u003eAs a baseline method, we implemented a greedy assignment heuristic with no cooperative optimisation. This Greedy method is a simple strategy wherein, at each step, it evaluates for each vehicle\u0026ndash;target pair the increase in the objective function if that target were assigned next to that vehicle, and then selects the vehicle\u0026ndash;target pair that yields the smallest cost increase. This assignment is repeated until all targets have been allocated, and finally each vehicle\u0026rsquo;s return leg to its depot is added to complete the routes. Because the objective function in our framework includes a weighting for collapsed points, the Greedy method does implicitly give some priority to collapsed sites (i.e. it considers not only distance but also the benefit of reaching a collapse sooner when choosing the next assignment). The Greedy method is computationally cheap and produces routes very quickly; however, due to its myopic, step-by-step nature, it often results in globally inefficient tours and uneven load distribution among vehicles (e.g. one vehicle may end up with a disproportionately long route).\u003c/p\u003e\u003cp\u003eFor evaluation, we compare the proposed method (Coop-GA) against the Greedy method in terms of key performance indicators such as the total time to complete coverage, the rapidity of collapse detection, and the balance of workload among vehicles. We also conduct comparisons on small-scale problems against brute-force optimal solutions, to verify how close the GA solutions are to the true optimum. Additionally, we simulate a scenario where some vehicles drop out during the mission and observe the effect of dynamic re-planning on route reconfiguration. As a visualisation, we generated a plot of all vehicle routes and survey points on the map (see Fig.\u0026nbsp;4) to illustrate the coverage and the area each vehicle is responsible for. Using the above setup, we perform a comprehensive evaluation of the framework\u0026rsquo;s operation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"4. Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Performance Comparison: Greedy vs Cooperative GA\u003c/h2\u003e\u003cp\u003eFirst, we compare the route assignment results produced by the Greedy heuristic and the cooperative GA (Coop-GA) method. For the standard scenario (9 vehicles and ~\u0026thinsp;70 target points, including 20 collapsed sites), Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarises the performance metrics of the plans obtained by both methods. Under the Greedy method, the route for the last vehicle became very long, and the overall completion time \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e was about 4.38 hours. In contrast, Coop-GA was able to reallocate tasks and shorten routes through genetic optimisation, reducing \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e to about 3.25 hours \u0026ndash; a reduction of roughly 1.13 hours (25%). The total sum of arrival times at all collapsed sites was also improved: from approximately 18.99 vehicle-hours under Greedy (the sum of times at which each collapsed site was reached by the vehicles) down to 15.99 hours with Coop-GA, a\u0026thinsp;~\u0026thinsp;16% reduction, indicating that collapsed building locations were incorporated earlier in the patrol order. Furthermore, the standard deviation of vehicle mission times (each vehicle\u0026rsquo;s total route duration) decreased from 1.48 hours with Greedy to 1.19 hours with Coop-GA, meaning the workload became more evenly distributed. In other words, under Greedy one particular vehicle ended up driving for a significantly longer time than the others, whereas Coop-GA produced a plan in which all vehicles finished their assignments in roughly equal times. Additionally, the overall objective function value \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:J\\)\u003c/span\u003e\u003c/span\u003e (as defined in Section \u003cspan refid=\"Sec6\" class=\"InternalRef\"\u003e2.3\u003c/span\u003e) improved from 472.1 for Greedy to 352.7 for Coop-GA, about a 25% improvement (lower is better). These results confirm that the proposed cooperative GA method outperforms the greedy assignment in terms of shorter completion time, quicker discovery of collapsed buildings, and more balanced resource use.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of performance metrics for the standard scenario (9 vehicles, ~\u0026thinsp;70 targets including 20 collapses).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e (h)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSum of collapse arrival times (h)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eStd. dev. of vehicle times (h)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eObjective \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:J\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGreedy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.38\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e18.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.48\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e472.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCoop-GA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.25\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e15.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e352.7\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn qualitative terms, the Greedy method\u0026rsquo;s distance-prioritised assignments caused a few faraway targets to all be left to one final vehicle, forcing that vehicle to make a large detour around almost the entire region. By contrast, Coop-GA\u0026rsquo;s genetic optimisation allowed task exchanges and route restructuring along the way, so far-off targets were split among multiple vehicles \u0026ndash; resulting in a plan in which \u0026ldquo;no target was left behind\u0026rdquo; and everyone shared the work. Some collapsed sites were also deferred until later in the Greedy plan, whereas Coop-GA, due to the collapse-priority term in the objective, tended to schedule collapsed-site visits earlier in the route order.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Near-Optimality and Computation Time\u003c/h2\u003e\u003cp\u003eNext, we evaluated how close the Coop-GA solutions are to the true optimal solution on smaller problem instances, and examined the computation cost. We limited the problem to 3 vehicles and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=6,7,8\\)\u003c/span\u003e\u003c/span\u003e targets, and performed an exhaustive search to find the optimal solution for each (since a brute-force search is only tractable at this small scale). We then compared the objective values of the Coop-GA solution and the Greedy solution against the optimum. The results showed that the Coop-GA\u0026rsquo;s objective value was within only a few percent of the exhaustive optimum, whereas the Greedy solution was substantially worse. For example, in the case of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=6\\)\u003c/span\u003e\u003c/span\u003e targets, if we normalise the exhaustive optimum\u0026rsquo;s objective value to 100%, the Coop-GA solution was about ~\u0026thinsp;102%, while the Greedy solution was around ~\u0026thinsp;120%. For \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=7\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=8\\)\u003c/span\u003e\u003c/span\u003e, the Coop-GA remained within ≲105% (near-optimal), whereas the Greedy solutions were significantly higher (in one example between ~\u0026thinsp;115\u0026ndash;140% of the optimum). In particular, when comparing the completion time \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, Greedy tends to be considerably larger than the optimal (because one vehicle\u0026rsquo;s route becomes excessively long), whereas Coop-GA, by suppressing the load of the slowest vehicle, achieved a \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e close to the optimal value. Also, regarding the speed of reaching the first collapse point, Coop-GA\u0026mdash;thanks to the collapse-weighting in the objective\u0026mdash;visits collapsed targets in an order close to the optimal strategy, whereas Greedy (which overly prioritises distance) often delays some collapses until later.\u003c/p\u003e\u003cp\u003eFrom a computation cost perspective, the Greedy algorithm decides assignments one task at a time, so the number of evaluations grows roughly on the order of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:O\\left(M\\times\\:K\\right)\\)\u003c/span\u003e\u003c/span\u003e (a quadratic function of number of vehicles \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:M\\)\u003c/span\u003e\u003c/span\u003e and targets \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K\\)\u003c/span\u003e\u003c/span\u003e ). On the other hand, the Coop-GA evaluates POP individuals per generation and runs for GEN generations, so its complexity is approximately \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:O\\left(\\text{POP}\\times\\:\\text{GEN}\\right)\\)\u003c/span\u003e\u003c/span\u003e. In practice, POP and GEN can be scaled roughly linearly with problem size; in our setting, POP\u0026thinsp;\u0026asymp;\u0026thinsp;30 and GEN\u0026thinsp;\u0026asymp;\u0026thinsp;60 were sufficient, and as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K\\)\u003c/span\u003e\u003c/span\u003e increased, the runtime scaled approximately linearly. Empirically, we found that even for a problem of 9 vehicles and 70 targets, the route optimisation could converge in on the order of seconds to at most tens of seconds. By contrast, exhaustive search suffers combinatorial explosion: even at \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=8\\)\u003c/span\u003e\u003c/span\u003e it had to evaluate over 1.8\u0026nbsp;million possibilities, and at \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=70\\)\u003c/span\u003e\u003c/span\u003e the solution space would be astronomically large (on the order of 10^110 combinations), making brute force infeasible. Therefore, the framework is practically justified in that it can obtain near-optimal solutions very quickly. With sufficient computing resources, one could further accelerate this by parallelising or distributing the GA computations, allowing even larger problem sizes to be handled efficiently.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eExhaustive optimum vs. Coop-GA vs. Greedy for small problems (3 vehicles \u0026times; \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=6,7,8\\)\u003c/span\u003e\u003c/span\u003e targets). (Exhaustive best is normalised to 100%.)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVehicles\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTargets \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eExhaustive Best (100%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eCoop-GA (% of optimum)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eGreedy (% of optimum)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eExhaustive search evaluations\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026asymp;\u0026nbsp;102%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026asymp;\u0026nbsp;120%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e20,160\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e≲\u0026nbsp;105% (near-optimal)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026gt;\u0026nbsp;Coop (e.g. 115\u0026ndash;140%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e181,440\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e≲\u0026nbsp;105% (near-optimal)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026gt;\u0026nbsp;Coop (e.g. 115\u0026ndash;140%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1,814,400\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Dynamic Re-planning Scenario\u003c/h2\u003e\u003cp\u003eIn the latter part of the simulation, we tested a scenario where a vehicle becomes incapacitated partway through the mission (to simulate a breakdown or loss of communication), and assessed the effect of dynamic task reallocation by Coop-GA. Specifically, at the 50% mark of the mission, we removed one vehicle, and invoked the Coop-GA re-optimisation to redistribute that vehicle\u0026rsquo;s remaining targets to the others. Under the Greedy (non-cooperative) approach, the area that had been assigned to the dropped vehicle initially remained unvisited, and eventually the other vehicles had to cover those targets sequentially on their own \u0026ndash; as a result, the overall completion time \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e was significantly extended. In contrast, with the Coop-GA approach, the dropout event was detected and GA was immediately re-run to reassign the remaining tasks of the lost vehicle to the others; as a result, the increase in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e was kept very small. In other words, the cooperative system quickly reallocated the workload, so that even with one vehicle missing the team\u0026rsquo;s delay in finishing was minimised. The qualitative comparison in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e also rated Coop-GA superior in terms of \u0026ldquo;scalability \u0026amp; fault tolerance,\u0026rdquo; noting that when a vehicle dropped out, the increase in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e was much smaller for Coop-GA, reflecting its advantageous re-planning capability.\u003c/p\u003e\u003cp\u003eTaken together, the results indicate that the proposed cooperative multi-vehicle system exhibits significantly higher performance than a static greedy heuristic. It enables rapid wide-area coverage and swift arrival at critical (collapsed) locations, allows efficient division of labour with no redundant overlap across multiple agencies, and also demonstrates better robustness to changing conditions.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Damage Detection Performance\u003c/h2\u003e\u003cp\u003eWe conducted an experiment (simulating the Noto Peninsula damage scenario) to evaluate the collapse classification performance of the proposed gNCDI\u0026thinsp;+\u0026thinsp;GA image analysis method. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e, we applied our model to features extracted from images labeled \u0026ldquo;damage present\u0026rdquo; and \u0026ldquo;no damage\u0026rdquo; respectively. The resulting confusion matrix of the classification was: TP\u0026thinsp;=\u0026thinsp;9, FP\u0026thinsp;=\u0026thinsp;2, TN\u0026thinsp;=\u0026thinsp;35, FN\u0026thinsp;=\u0026thinsp;1. In this test, the method achieved a Precision\u0026thinsp;=\u0026thinsp;0.8182 and Recall\u0026thinsp;=\u0026thinsp;0.9000, yielding an overall F1 score\u0026thinsp;=\u0026thinsp;0.8571. In other words, on this limited validation set the method detected 90% of collapsed buildings without missing any, while keeping false alarms to about 18%. In quantitative terms, both precision and recall improved compared to a conventional method that uses a simple RI threshold, and a significant improvement in F1 was observed. For example, cases of collapsed houses covered by blue tarpaulins (low red component), or cases with only wall cracking and little rubble, were difficult for RI to detect, but the proposed method correctly identified the damage by leveraging colour-difference information. The sample detection results in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e also show that our gNCDI\u0026thinsp;+\u0026thinsp;GA approach can pick up damage sites that the RI-based method missed. These findings indicate that the proposed method enables high-accuracy damage judgement even when only limited image data are available, and that qualitatively it expanded the range of damage patterns that can be detected, enhancing its practical utility in the field.\u003c/p\u003e\u003cp\u003eConsidering both qualitative and quantitative evaluation results, we gained confidence that this framework can automatically generate a reliable damage map within a few hours of an earthquake. In addition to confirming its rapidity and high spatial resolution advantage over traditional satellite or aerial imagery approaches, we also identified practical operational benefits of our method, such as inherent data privacy (using ground-level imagery rather than broad aerial surveillance) and a reduction in false detections caused by red-colored artifacts. These advantages, as well as remaining challenges, are discussed in the next section.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Advantages and Limitations\u003c/h2\u003e\u003cp\u003eAdvantages\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eBalance of speed and accuracy: Whereas satellite-derived damage maps are often available only the day after an event or later, our framework can produce a detailed, ground-level damage map within a few hours. In simulation, 9 vehicles achieved full-area coverage in a little over 3 hours, which is less than half the time required by a single-vehicle operation (approximately inversely proportional to the number of vehicles).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHigh collapse-detection performance: For EMS-98 D5 (collapse) detection, the gNCDI-based approach can be expected to achieve\u0026thinsp;\u0026asymp;\u0026thinsp;90% recall, with false detections held to \u0026asymp;\u0026thinsp;20% or less (18% in our evaluation). Because the thresholding prioritizes recall to minimize misses, some false positives do occur; however, since missing a collapse is more critical, this trade-off is acceptable in practice.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eOperational benefits: A cloud-based architecture standardizes JSON sharing across agencies and maintains a single, real-time damage map. By integrating learning-free, lightweight image analysis (gNCDI) with GA-based vehicle route optimization, the framework enhances immediacy, scalability, and field deployability.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eLimitations (from the perspective of gNCDI-based image classification)\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eSensitivity to damage grade: As an RGB color-difference index, gNCDI is highly reliable for severe collapses where exposed timber/soil is prominent, but it tends to under-detect partial/moderate damage (EMS-98 D3/D4, e.g., wall cracking) where color changes are subtle.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eImaging-condition dependence: Performance may degrade under nighttime/deep shadow, mixed lighting/auto white balance, motion blur, heavy dust/smoke, or color casts. While false positives from uniformly red structures (e.g., red roof tiles or painted walls) are reduced compared to simple RI, residual confounders remain when pigments, banners, rust, or background soils exhibit similar chromatic characteristics.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDomain shift and threshold tuning: Differences in camera models, compression, and weather can alter color reproduction, causing the optimal threshold to drift. Per-scene calibration (using a simple reference image or a few samples) is recommended.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eOcclusion and visibility constraints: Vegetation or adjacent structures may occlude rubble from ground cameras. This can be mitigated by complementing ground surveys with UAV/helicopter observations and by multi-pass coverage from varied viewpoints.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Redness Index Sensitivity and Calibration\u003c/h2\u003e\u003cp\u003eThe sensitivity of the Redness Index (and by extension gNCDI) can vary with local building materials and imaging conditions. For example, in regions with many wooden houses, a collapse results in conspicuous brownish-red timber and earth dust, yielding a strong RI signal; in contrast, in modern urban areas dominated by reinforced concrete, a collapse produces mostly gray debris and RI may respond weakly. Therefore, it is important to perform pre-calibration to set appropriate thresholds or colour corrections tailored to the region. Moreover, since imagery in bright midday conditions will have different colour characteristics than images taken in the evening or at night, applying brightness or colour temperature adjustments as needed can help maintain detection accuracy. In our experiments, we found that by using gNCDI (which generalises colour differences) instead of the simple redness index, we achieved better recall, precision, and F1 scores across varying conditions. This suggests that the generalised colour-difference approach is inherently more robust to differences in lighting and scene composition.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e5.3 Load Balancing and Redundancy Control\u003c/h2\u003e\u003cp\u003eA major outcome of introducing the cooperative GA was the significant equalisation of task load among the multiple vehicles. Under Greedy assignment, it was common for one vehicle to end up with a concentration of targets at the end, leading to a large standard deviation in travel times [14]. This not only overburdens that particular vehicle, but also means the entire team\u0026rsquo;s response is prolonged while waiting for that single slowest vehicle to finish. Coop-GA, by including a deviation penalty in its objective, adjusted the assignment so that each vehicle\u0026rsquo;s distance/time is roughly even, yielding a plan where everyone can finish in step together. This is desirable also from the perspective of distributing human workload on the ground (more equitable fatigue levels).\u003c/p\u003e\u003cp\u003eRegarding redundancy (overlapping coverage), by scenario design each target is only visited once, so the theoretical direct overlap is zero. However, inefficient routing can lead to indirect redundancy in the form of unnecessary detours. In the Greedy method, the order in which vehicles are assigned targets sometimes caused some vehicles to take long detours (effectively a kind of redundant travel). Coop-GA, through inter-vehicle task swaps and intra-route 2-opt optimisation, was able to compress this wasted travel. As a result, the proportion of redundant distance out of total distance driven was slightly reduced, indicating that cooperative GA is advantageous in curbing waste of transportation resources as well.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e5.4 Behaviour under Vehicle Dropout and Re-planning\u003c/h2\u003e\u003cp\u003eAs noted above, the Coop-GA approach possesses adaptability to dynamic changes such as vehicle dropout or new task arrival. A vehicle dropout (failure or loss of communication) is a very realistic possibility in the field, and traditional fixed plans cannot smoothly reallocate the remaining tasks. With Greedy assignment, the plan is decided once at the start and lacks flexibility, so if a dropout occurs, an additional delay is incurred while other vehicles eventually cover the orphaned tasks. In contrast, Coop-GA can immediately re-run its optimisation upon detecting a dropout, reassigning the lost vehicle\u0026rsquo;s tasks to the remaining vehicles such that the increase in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{T}_{\\text{m}\\text{a}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e is minimised. This difference arises from the presence (or absence) of a real-time route re-optimisation capability, which is a major strength of our framework. In actual field operations, there are limitations in communication infrastructure and computing resources, but it is feasible to perform periodic re-computations on the order of every few seconds. Having multiple vehicles flexibly adjust their coverage areas in cooperation greatly enhances robustness under the uncertain and evolving conditions of a disaster.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003e5.5 Integration with Other Methods and Extensibility\u003c/h2\u003e\u003cp\u003eThe proposed framework can function complementarily with existing damage assessment methods. For example, combining it with broad-area aerial imaging (e.g. from reconnaissance aircraft) or detailed assessments by structural experts would enable a more reliable and comprehensive understanding of damage. The damage maps automatically generated by our system can be integrated as a real-time layer into the common operational picture at disaster response headquarters, aiding decisions on where to deploy resources. Cross-referencing it with next-day satellite-based macro-scale damage maps could help to fill in any missed spots and to verify potential false positives. The framework itself is also highly generalisable and could be applied to other types of disasters such as heavy rain, typhoons, or floods. By swapping in different indices or sensor inputs appropriate to the disaster (for instance, a thermal infrared index for detecting burnt buildings in wildfires), one can utilise the ground-vehicle platform for rapid damage assessment across a wide range of scenarios. Thus, the approach offers broad applicability by adapting its indices and sensor payloads to different disaster types.\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eAcknowledging the operational reality that, in the immediate aftermath of a disaster, there is no time to build or retrain deep learning models, this study presented a comprehensive framework that achieves wide-area, real-time damage assessment. The framework integrates collapsed-building detection from in-vehicle imagery using gNCDI, multi-agent route optimization via a Genetic Algorithm (GA), and inter-agency coordination through cloud-based JSON distribution. We implemented and validated the framework using both simulation and real-world footage. In a Noto Peninsula\u0026ndash;style evaluation, coordination of nine vehicles indicated the potential to generate a wide-area damage map within a few hours.\u003c/p\u003e\u003cp\u003eThe gNCDI-based collapse detection achieved Precision\u0026thinsp;=\u0026thinsp;0.8182, Recall\u0026thinsp;=\u0026thinsp;0.9000, and F1\u0026thinsp;=\u0026thinsp;0.8571 (TP\u0026thinsp;=\u0026thinsp;9, FP\u0026thinsp;=\u0026thinsp;2, TN\u0026thinsp;=\u0026thinsp;35, FN\u0026thinsp;=\u0026thinsp;1), outperforming a simple RI-threshold method quantitatively. Compared with a greedy assignment, the GA produced more efficient patrol plans, shortening the overall completion time T\u003csub\u003emax\u003c/sub\u003e by about 25% while balancing workload, and converged to a near-optimal solution within seconds despite a search space on the order of 10\u003csup\u003e110\u003c/sup\u003e combinations. Moreover, even under dynamic changes such as vehicle dropout or road closures, rapid reallocation maintained full coverage; false detections remained at a level suppressible by subsequent human verification.\u003c/p\u003e\u003cp\u003eTaken together, the framework is superior to conventional satellite/aerial approaches in both speed and spatial resolution and can effectively support early rescue and recovery. Looking ahead, we aim to evolve it into a standard, multi-hazard platform applicable to complex disasters (e.g., earthquakes and typhoons) that contributes to damage reduction and the protection of human life. The authors sincerely hope this work will help accelerate life-saving efforts.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eG\u0026oacute;mez-Chova, Luis, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Multimodal classification of remote sensing images: A review and future directions.\u0026rdquo; \u003cem\u003eProceedings of the IEEE\u003c/em\u003e 103.9 (2015): 1560\u0026ndash;1584.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTu, Jihui, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Automatic building damage detection method using high-resolution remote sensing images and 3D GIS model.\u0026rdquo; \u003cem\u003eISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences\u003c/em\u003e 3 (2016): 43\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTilon, Sofia, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Post-disaster building damage detection from earth observation imagery using unsupervised and transferable anomaly detecting generative adversarial networks.\u0026rdquo; \u003cem\u003eRemote Sensing\u003c/em\u003e 12.24 (2020): 4193.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMonfort, Daniel, Caterina Negulescu, and Myriam Belvaux. \u0026ldquo;Remote sensing vs. field survey data in a post-earthquake context: Potentialities and limits of damaged building assessment datasets.\u0026rdquo; \u003cem\u003eRemote Sensing Applications: Society and Environment\u003c/em\u003e 14 (2019): 46\u0026ndash;59.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSuppasri, Anawat, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;The 2024 Noto Peninsula earthquake: Preliminary observations and lessons to be learned.\u0026rdquo; \u003cem\u003eInternational Journal of Disaster Risk Reduction\u003c/em\u003e 110 (2024): 104611.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e2024 Noto earthquake bosaiXview. https://xview.bosai.go.jp/view/index.html?appid=41a77b3dcf3846029206b86107877780 (accessed 16 Jan 2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMizui, Y., Fujiwara, H. \u003cem\u003eEstimate the amount of disaster waste disposal work using in-vehicle camera images \u0026ndash; a case study in Hitoyoshi City, Kumamoto Prefecture. Journal of Disaster Research\u003c/em\u003e, 16(7), 2021, pp. 1061\u0026ndash;1073. https://doi.org/10.20965/jdr.2021.p1061.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVan der Meer, Freek. \u0026ldquo;Remote-sensing image analysis and geostatistics.\u0026rdquo; \u003cem\u003eInternational Journal of Remote Sensing\u003c/em\u003e 33.18 (2012): 5644\u0026ndash;5676.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa, Lei, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Deep learning in remote sensing applications: A meta-analysis and review.\u0026rdquo; \u003cem\u003eISPRS Journal of Photogrammetry and Remote Sensing\u003c/em\u003e 152 (2019): 166\u0026ndash;177.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEscadafal, R., Belghit, A., \u0026amp; Ben-Moussa, A. Indices spectraux pour la t\u0026eacute;l\u0026eacute;d\u0026eacute;tection de la d\u0026eacute;gradation des milieux naturels en Tunisie aride, \u003cem\u003ein\u003c/em\u003e: Guyot G. (ed.), \u003cem\u003eMesures physiques et signatures en t\u0026eacute;l\u0026eacute;d\u0026eacute;tection: Sixi\u0026egrave;me symposium international; actes; 17\u0026ndash;21 Janvier 1994, Val d\u0026rsquo;Is\u0026egrave;re, France: Proceedings\u003c/em\u003e, CNES (Centre National d\u0026rsquo;Etudes Spatiales), Paris, 1994, pp. 253\u0026ndash;259.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEscadafal, R., \u0026amp; Huete, A. \u0026Eacute;tude des propri\u0026eacute;t\u0026eacute;s spectrales des sols arides appliqu\u0026eacute;e \u0026agrave; l\u0026rsquo;am\u0026eacute;lioration des indices de v\u0026eacute;g\u0026eacute;tation obtenus par t\u0026eacute;l\u0026eacute;d\u0026eacute;tection [Improvement in remote sensing of low vegetation cover in arid regions by correcting vegetation indices for soil \u0026ldquo;noise\u0026rdquo;]. \u003cem\u003eComptes Rendus de l\u0026rsquo;Acad\u0026eacute;mie des Sciences, S\u0026eacute;rie II\u003c/em\u003e 312(11), 1991, pp. 1385\u0026ndash;1391.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShiraishi, H., \u0026amp; Usuda, Y. (2025). \u003cem\u003eReal-Time Building-Damage-Extraction Technology from Ground-Based Video Footage Using Normalized Difference Red/Green Redness Index\u003c/em\u003e. Geomatics and Environmental Engineering, 19(1), 143\u0026ndash;159. https://doi.org/10.7494/geom.2025.19.1.143\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTaniguchi, E., Yamada, T., \u0026amp; Kakimoto, Y. Probabilistic vehicle routing and scheduling with variable travel times. \u003cem\u003eIFAC Proceedings\u003c/em\u003e, 33(9), 2000, pp. 33\u0026ndash;38. https://doi.org/10.1016/S1474-6670(17)38119-3.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKefi, M., \u0026amp; Ghedira, K. \u0026ldquo;A multi-agent model for the Vehicle Routing Problem with Time Windows.\u0026rdquo; \u003cem\u003eWIT Transactions on The Built Environment\u003c/em\u003e 75 (2004).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZidi, Kamel, \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Distributed genetic algorithm for disaster relief planning.\u0026rdquo; \u003cem\u003eInternational Journal of Computers Communications \u0026amp; Control\u003c/em\u003e 8.5 (2013): 769\u0026ndash;783.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"University of Tsukuba","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Rapid Earthquake Damage Estimation, Ground-Level RGB Image Analysis (gNCDI), Cooperative Multi-Vehicle Routing (Genetic Algorithm), Inter-Agency Coordination (ICS), Real-Time Disaster Mapping","lastPublishedDoi":"10.21203/rs.3.rs-7390282/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7390282/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eQuickly and accurately evaluating building damage after a major earthquake is essential for effective emergency response. We propose a practical framework that integrates camera-equipped ground vehicles deployed by multiple agencies (such as the Self-Defense Forces, police, and fire services) with a centralized command system to perform real-time post-earthquake damage mapping. The system combines an RGB-based damage detection technique (gNCDI), which generalises the simple Redness Index (RI) originally developed for vegetation analysis, with a Genetic Algorithm (GA) to optimise the patrol routes of multiple vehicles. Using colour-based inference, collapsed buildings are rapidly identified from ground-level images by detecting the spectral signatures of exposed timber and soil debris, while the GA efficiently allocates routes to each vehicle to maximise coverage and minimise response time. A cloud-based architecture standardises and shares geotagged damage reports in real time using a JSON format across all responding agencies. We present the system design, implementation details, and evaluation protocol based on a disaster scenario simulation for the Noto Peninsula region in Japan. In our evaluation, the proposed approach achieved a high overall classification accuracy (F1 score\u0026thinsp;\u0026asymp;\u0026thinsp;0.86), detecting 90% of collapsed buildings with only\u0026thinsp;~\u0026thinsp;18% false alarms. At the same time, the cooperative vehicle-routing strategy significantly improved survey efficiency, shortening total mission completion time by around 25% compared to a greedy baseline. Furthermore, we discuss practical issues including the speed and resolution advantages over traditional satellite or aerial assessments, data privacy considerations, false detections, and the need for human verification of results. Overall, this study demonstrates a feasible multi-vehicle, multi-agency approach for rapid earthquake damage estimation aimed at accelerating life-saving rescue operations and optimising resource allocation.\u003c/p\u003e","manuscriptTitle":"A Practical Framework for Rapid Earthquake Damage Estimation through Multi-Vehicle Coordination and Inter-Agency Collaboration: Integrating Genetic Algorithms with RGB-Based Image Analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-20 06:48:14","doi":"10.21203/rs.3.rs-7390282/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e76f3b7c-cb8e-49ce-88e5-d41ca0441c82","owner":[],"postedDate":"August 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":53256452,"name":"Computer Architecture and Engineering"}],"tags":[],"updatedAt":"2025-11-20T18:45:23+00:00","versionOfRecord":{"articleIdentity":"rs-7390282","link":"https://doi.org/10.1016/j.nhres.2025.11.003","journal":{"identity":"natural-hazards-research","isVorOnly":true,"title":"Natural Hazards Research"},"publishedOn":"2025-11-19 00:00:00","publishedOnDateReadable":"November 19th, 2025"},"versionCreatedAt":"2025-08-20 06:48:14","video":"","vorDoi":"10.1016/j.nhres.2025.11.003","vorDoiUrl":"https://doi.org/10.1016/j.nhres.2025.11.003","workflowStages":[]},"version":"v1","identity":"rs-7390282","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7390282","identity":"rs-7390282","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}