A Hybrid Decision Support Framework Integrating Combined Probabilistic Spatial Modeling with Large Language Models for Post-Earthquake Search and Rescue

Abstract

This study describes the design and implementation of a hybrid decision support framework for post-earthquake urban Search and Rescue (SAR) prioritization, which combines two probabilistic spatial models with a Large Language Model (LLM). This approach examines the combination of two distance decay models: truncated negative exponential and lognormal models with the aim to transform discrete geolocated incident reports into a probabilistic priority surface. High-priority hotspots are identified using thresholding and spatial clustering. The proposed framework includes the use of the OpenAI model in its open source version for generating structured SAR recommendations. This approach is assessed using a synthetic dataset indicating post-earthquake locations in Mexico City. In addition, a sensitivity analysis was carried out to show stability in hotspot ranking. Our preliminary results indicate that the recommendations generated by the LLM match the hotspot scores. Thus, our proposed framework provides a suitable integration among geospatial modeling and LLM features for reasoning in urban disaster decision support.

1. Introduction

Earthquakes are natural disasters that generate conditions which play an important role in shaping emergency response strategies; the presence of various features such as structural damage, high population density, and information overload during the early hours following a disaster event makes earthquakes an important case of study.

In high-density urban areas such as Mexico City, post-earthquake search and rescue (SAR) operations are carried out under conditions of uncertainty and resource constraints. In this context, decisions regarding the prioritization of search and rescue teams establish a critical decision-making challenge with direct implications for survival outcomes [1,2].

Recent developments have highlighted the need to integrate artificial intelligence into a wide range of human activities, specifically those that require well informed decision making under time pressure such as natural disasters [3,4].

It is important to note that natural disaster management has been a major area of research worldwide [5], mainly in regions with a significant history of this sort of event, such as Tokyo and Mexico City [6,7]. These major cities have experienced social and material damage due to earthquakes. Thus, there is an urgent need to address the safety and public health challenges in terms of seismic disasters [8].

One of the most important post-event assessments of the 19 September 2017 earthquake in Mexico City showed that structural damage and building collapses were not uniformly distributed [9]. Structural analyses showed that the earthquake tended to be clustered rather than spatially uniform. This pattern highlights the need for spatial strategies to support the prioritization of search and rescue operations [10]. However, much uncertainty still exists regarding how to manage disaster response in post-event situations; for instance, in relation to resource deployment, rapid response coordination, and the identification of priority rescue areas [11].

So far, very little attention has been paid to operational decision-making processes [12]. In this regard, probabilistic spatial decay models provide a mechanism for transforming discrete incident reports into continuous priority surfaces suitable for operational decision support [13]. In post-earthquake situations, the spatial influence of damage reports is rarely uniform [14].

This study proposes a combined approach based on the following decay models: the truncated negative exponential model (TNE) and the lognormal decay model (LN). On the one hand, the TNE describes a systematic distance-based decrease while limiting influence within an operational radius. On the other hand, the LN model adapts to asymmetric and irregular propagation patterns, which are frequently observed in urban seismic damage distributions.

Artificial intelligence integration into natural disaster assessment is a major area of interest within the field of crisis management where traditional techniques such as machine learning have previously been used to address challenges related to real-time earthquake detection and magnitude estimation [15,16]. Large language models have emerged as an interesting approach for contextual understanding and decision making; moreover, LLMs are able to analyze, classify and gain knowledge from different digital sources such as social networks [17,18].

Recent disaster management research suggests that reasoning approaches based on large language models may perform potentially well in complex disaster scenarios [19,20]. The integration of probabilistic spatial models with recommendations generated by LLMs remains an open challenge, specifically in situations where decision makers need to synthesize incoming damage reports, medical needs, and on the ground rescue operations.

This study proposes a hybrid decision support framework integrating two probabilistic models with LLM for post-earthquake search and rescue operations in Mexico City [21]. The aim of this research is to show that by using a combined probabilistic spatial model with a decision support layer, SAR operations are able to be enhanced in the immediate hours following an earthquake [22].

Our proposed hybrid framework intends to determine the extent to which a combined probabilistic spatial model can identify high-priority SAR hotspots [23,24]. In this context, our combined probabilistic spatial model is assessed through a synthetic earthquake scenario for Mexico City. The use of a synthetic dataset permits us to carry out sensitivity testing while preserving realism in terms of spatial distribution, incident density, and severity levels [25]. Thus, this study addresses the following research questions.

RQ1: To what extent can a combined probabilistic spatial model reliably identify high-priority SAR hotspots in post-earthquake urban environments such as Mexico City?

RQ2: How can locally deployed large language models enhance the operational effectiveness of probabilistic SAR hotspot detection without compromising reliability, explainability, or privacy?

It is important to clarify the scope of the present study. In this case, the main goal is to develop a hybrid decision support framework for ranking SAR operations based on geolocated incident reports during the early stage of a post-earthquake situation. It is important to note that the LLM recommendations are mainly used to complement, rather than replace, structural damage assessment methodologies commonly used in urban seismic risk analysis.

To the best of our knowledge, no previous study has integrated probabilistic spatial hotspot detection with a local LLM for SAR operational prioritization. In summary, the main contributions of this study are the following:

• A combined probabilistic spatial modeling approach for SAR prioritization.
• An LLM-assisted decision support layer, where a locally deployed large language model generates operational recommendations.
• A hybrid decision support framework for disaster response.

2. Materials and Methods

2.1. Earthquake Scenario Using Synthetic Dataset

Mexico City is placed within a seismically active zone shaped by the interaction between the Cocos and North American tectonic plates, which results in a high exposure to ground motion events [10]. These geographic conditions make Mexico City a challenging environment for post-earthquake SAR operations.

Due to the limited availability of real-time data during disaster response, this study uses a synthetic dataset. This dataset was intentionally designed to show realistic post-earthquake reporting patterns observed during the 2017 in Mexico City. Each record corresponds to an individual report and includes date and time information, report type, building type, damage level, estimated number of affected persons, and geographic coordinates in terms of latitude and longitude.

Although the dataset used in this study is synthetic, it was designed to approximate spatial reporting patterns observed after major urban earthquakes. This synthetic dataset therefore incorporates heterogeneous spatial clusters, varying damage levels, and differences in the number of reported affected individuals. Nevertheless, synthetic data cannot fully emulate a real-world emergency scenario.

The dataset only considers the spatial distribution of incidents concentrated in urban areas that have been historically associated with higher levels of structural damage and building collapse. The dataset used in this study is provided in Appendix A for reproducibility and transparency.

All incident records were adjusted into spatial objects for geostatistical analysis. This research uses a regular spatial grid across the study area to support raster modeling, allowing incident influence to be assessed over the entire urban domain rather than at isolated locations. The rasterisation step was selected for its reliability in probabilistic hotspot identification; the gain of this approach is that it supports spatial aggregation, neighborhood analysis, and the identification of contiguous high-priority regions.

Although rasterisation was selected in this research for its simplicity and compatibility with continuous spatial modeling, alternative approaches could be considered, e.g., kernel density estimation methods may be used to generate smooth spatial intensity surfaces directly from incident points [26]. Vector clustering approaches, such as spatial clustering algorithms or Voronoi tessellations may be applied to identify areas of incident concentration without discretizing the study area into regular cells [27].

2.2. Probabilistic Spatial Models

2.2.1. Truncated Negative Exponential Model

Given the set of geolocated incident reports $\mathcal{E}$:

$$\mathcal{E}={\left\{e_i=\left(x_i,t_i\right)\right\}}_{i=1}^N,$$

where $x_i\in {\mathbb{R}}^2$ denotes the spatial location of incident $i$, $t_i$ represents the reporting time, and $N$ is the total number of reported incidents.

$$P\left(x\right):{\mathbb{R}}^2\to \left[\mathrm{0,1}\right],$$

such that higher values indicate areas of greater SAR priority. Thus, a distance decay model is introduced to denote the way in which the influence of each incident decreases with distance [28]. Formally, for each incident $e_i$, the truncated negative exponential (TNE) spatial influence at location $x\in {\mathbb{R}}^2$ is defined as a conditional function:

$$f_i^{TNE}\left(x\right)=\left\{\begin{array}{c}\lambda \mathit{exp}\left(-‖x-x_i‖\right), ‖x-x_i‖\le d_{max}\\ 0, ‖x-x_i‖>d_{max}\end{array}\right.$$

where $\lambda >0$ controls the rate of decay, $\parallel x-x_i\parallel$ indicates the Euclidean distance between location $x$ and incident location $x_i$, and $d_{max}$ corresponds to the maximum distance at which an incident remains operationally relevant; beyond this threshold, its influence is ignored. The TNE priority surface is then computed as follows:

$$P_{TNE}\left(x\right)=\sum _{i=1}^N{f}_i^{TNE}\left(x\right),$$

which yields a continuous spatial field over ${\mathbb{R}}^2$. The truncation parameter makes the model more realistic by excluding effects that are beyond a meaningful spatial range. Even though the TNE model assumes that impact decreases steadily with distance, the urban damage may propagate in spatially varied patterns. In such cases, SAR operations might be relevant in surrounding areas rather than precisely at the incident location. This point is important, as it indicates that the priority hotspot will not necessarily be located at the damaged building; individuals may instead move toward safe zones or shelters.

2.2.2. Lognormal Decay Model

Similarly, this study uses a lognormal decay function (LN) to show how the impact changes over distance. The advantage of this approach is that in a post-earthquake situation this decay model allows us to observe spatial variability, i.e., under real circumstances the damage does not always decrease smoothly; instead, it may exhibit spatial heterogeneity [29].

Formally, for each incident $e_i=(x_i,t_i)$, the LN influence at location $x\in {\mathbb{R}}^2$ is defined as:

$$f_i^{LN}\left(x\right)={\displaystyle \frac{1}{‖x-x_i‖\sigma \sqrt{2\pi }}}\mathit{exp}\left(-{\displaystyle \frac{{\left(\mathit{ln}‖x-x_i‖-\mu \right)}^2}{2{\sigma }^2}}\right), ‖x-x_i‖>0,$$

where $\mathsf{\mu }\in \mathbb{R}$ denotes the logarithmic location parameter and $\mathsf{\sigma }>\mathbf{0}$ controls the dispersion of the distribution. The LN priority surface is then computed as follows:

$${\mathit{P}}_{\mathit{L}\mathit{N}}\left(\mathit{x}\right)={\displaystyle \sum _{\mathit{i}=\mathbf{1}}^{\mathit{N}}}{\mathit{f}}_{\mathit{i}}^{\mathit{L}\mathit{N}}\left(\mathit{x}\right),$$

which yields a continuous spatial field over ${\mathbb{R}}^{\mathbf{2}}$.

2.2.3. Combined Probabilistic Spatial Model

Based on the two previously mentioned models, a combined probabilistic spatial modeling approach was proposed to address post-earthquake disaster management. In this context, due to the magnitudes of ${\mathit{P}}_{\mathbf{T}\mathbf{N}\mathbf{E}}\left(\mathit{x}\right)$ and ${\mathit{P}}_{\mathbf{L}\mathbf{N}}\left(\mathit{x}\right)$ may differ, both functions are scaled to the same spatial domain to ensure standardization. Our combined probabilistic spatial model is as follows:

$${\mathit{P}}_{\mathit{T}\mathit{N}\mathit{E}}^{\mathit{n}\mathit{o}\mathit{r}\mathit{m}}\left(\mathit{x}\right)={\displaystyle \frac{{\mathit{P}}_{\mathit{T}\mathit{N}\mathit{E}}\left(\mathit{x}\right)}{\mathit{m}\mathit{a}{\mathit{x}}_{\mathit{x}\in {\mathbb{R}}^{\mathbf{2}}}{\mathit{P}}_{\mathit{T}\mathit{N}\mathit{E}}\left(\mathit{x}\right)}},{\mathit{P}}_{\mathit{L}\mathit{N}}^{\mathit{n}\mathit{o}\mathit{r}\mathit{m}}\left(\mathit{x}\right)={\displaystyle \frac{{\mathit{P}}_{\mathit{L}\mathit{N}}\left(\mathit{x}\right)}{\mathit{m}\mathit{a}{\mathit{x}}_{\mathit{x}\in {\mathbb{R}}^{\mathbf{2}}}{\mathit{P}}_{\mathit{L}\mathit{N}}\left(\mathit{x}\right)}}.$$

A weighted average of both models provides the resultant SAR priority surface ${\mathit{P}}_{\mathrm{SAR}}\left(\mathit{x}\right):{\mathbb{R}}^{\mathbf{2}}\to \left[\mathbf{0},\mathbf{1}\right]$.

$${\mathit{P}}_{\mathit{S}\mathit{A}\mathit{R}}\left(\mathit{x}\right)=\mathit{\alpha }{\mathit{P}}_{\mathit{T}\mathit{N}\mathit{E}}^{\mathit{n}\mathit{o}\mathit{r}\mathit{m}}\left(\mathit{x}\right)+\mathit{\beta }{\mathit{P}}_{\mathit{L}\mathit{N}}^{\mathit{n}\mathit{o}\mathit{r}\mathit{m}}\left(\mathit{x}\right),$$

where $\mathit{\alpha },\mathit{\beta }\ge \mathbf{0}$ and $\mathit{\alpha }+\mathit{\beta }=\mathbf{1}$. This weighted combination allows the combined model to balance the distance dependent decay functions with the irregular patterns of real earthquake damage. It should be noted that the weighting parameters $\mathit{\alpha },\mathit{\beta }$ may be calibrated according to empirical tuning such as expert judgment, urban density and infrastructural weakness. In this context, the weighting parameters α and β determine the contribution of the two decay models in the combined probabilistic surface. Higher values of α increase the influence of the TNE component, leading to more spatially concentrated regions around incident clusters. On the contrary, larger contributions from the LN component allow the model to capture irregular spatial propagation patterns that may arise from heterogeneous damage distributions.

On the other hand, hotspots $H\left(x\right)$ are identified by applying a statistical threshold over the combined probabilistic surface $P_{\mathrm{SAR}}\left(x\right)$. Let $\tau$ denote as a threshold such as the 95th percentile. A binary mask is defined as follows:

$$H\left(x\right)=\left\{\begin{array}{c}1, P_{SAR}\left(x\right)\ge \tau \\ 0, P_{SAR}\left(x\right)<\tau \end{array}\right.$$

Formally, a hotspot region ${\mathcal{R}}_{\mathit{k}}\subset {\mathbb{R}}^{\mathbf{2}}$ is defined as a maximal connected subset of locations for which $\mathit{H}\left(\mathit{x}\right)=\mathbf{1}$. For each detected region ${\mathit{R}}_{\mathit{k}}$, summary indicators are computed to support operational ranking. The maximum priority value within the region is defined as:

$${\mathit{S}}_{\mathit{k}}=\underset{\mathit{x}\in {\mathit{R}}_{\mathit{k}}}{\mathbf{max}}{\mathit{P}}_{\mathit{S}\mathit{A}\mathit{R}}\left(\mathit{x}\right),$$

while the spatial extent is computed as:

$$A_k=\left|R_k\right|a_{cell},$$

where $\mid R_k\mid$ denotes the number of raster cells in region $R_k$ and $a_{cell}$ indicates the area of a single cell. The centroid $c_k\in {\mathbb{R}}^2$ is obtained as the geometric means of all spatial coordinates within the region $R_k$,

$$c_k={\displaystyle \frac{1}{\left|R_k\right|}}\sum _{x\in R_k}x$$

Regions are ranked according to descending values of $S_k$, and only the top $K$ regions are retained for operational consideration; in this study, $K=3$ is used to simulate realistic urban earthquake scenarios.

The parameter $K$ stands for the number of hotspot regions selected for operational prioritization. In practical SAR scenarios, the value of $K$ may vary depending on operational constraints such as the number of available response teams, the spatial scale of the disaster, and the density of incident reports. Lower values of $K$ may be appropriate in situations with limited resources, where only the most critical areas can be attended immediately.

In general terms, a spatial decay function represents the assumption that the influence of an incident report decreases with distance from its location. Points located closer to reported incidents therefore receive higher priority scores, while more distant areas receive lower values. This behavior allows the model to transform discrete observations into a continuous spatial representation of potential SAR priority zones.

The weighting parameters α and β regulate the relative contribution of the TNE and LN decay functions. Similar weighting strategies are usually used in spatial interaction models and geographic profiling approaches [30,31]. In this study, the influence of these parameters was explored through sensitivity analysis in Section Hotspot Stability Under Alpha Variationusing values α = 0.5, 0.6, and 0.7. The results indicate that hotspot rankings remain relatively stable across this range.

2.3. LLM-Assisted Decision Support Layer

While our proposed combined probabilistic spatial framework provides a quantitative ranking surface, operational search and rescue tasks need contextual interpretation and tactical recommendations. The current study enables the integration of a layer based on LLMs that interact with data derived from the combined probabilistic spatial model.

LLMs are able to process text using units known as tokens, which represent fragments of words or characters used internally by the model to encode language inputs and outputs. The maximum number of tokens that can be processed at once is referred to as the context length window, which determines the amount of information the model can consider when generating a response [4]. In addition, the generation process is influenced by a parameter known as temperature, which controls the level of randomness in the produced text. Lower temperature values typically produce more deterministic and conservative responses, whereas higher values allow more variation in the generated outputs.

In order to assess whether an LLM was able to provide reliable recommendations based on the combined probabilistic surface $P_{\mathrm{SAR}}\left(x\right)$, we used the OpenAI gpt-oss-20b model [22]; this open source model was deployed locally using LM Studio version 0.4.2, which provides a REST compatible local server interface [32]. This architecture ensures full data privacy and avoids external API dependency.

The LLM allows a structured prompt submission and a controlled output generation. The prompt included the following data: hotspot rank, priority score, spatial extent, and contextual earthquake information. Thus, this prompt enables the LLM to analyze emergency situations, prioritize critical information, and generate recommendations aligned with established emergency protocols.

Figure 1 shows an overview of the visual integration between the combined probabilistic spatial model and the decision support layer. It can be seen from the figure that the raster map displays the spatial priority field $P_{\mathrm{SAR}}\left(x\right)$, in yellow tones, while hotspot centroids (red pines) activate the operational recommendations (pop-up message) generated by the OpenAI gpt-oss-20b local model via LM Studio.

To improve transparency and reduce hallucinations, the prompt template used in this study was designed as a constrained instruction format. The model was required to generate responses using the following semantic categories: priority level, team allocation, recommended actions, risk identification, and response time. These categories were explicitly specified in the main prompt template (see Figure 2). This prompt design allowed us to reduce the likelihood of generating speculative responses and encourages the model to produce recommendations aligned with established urban SAR operational practices. It should be noted that this research does not include a formal expert assessment.

This prompt enables the local LLM to analyze emergency situations on the resultant SAR priority surface $P_{\mathrm{SAR}}\left(x\right)$. This surface was rendered as a color graded heatmap over a geographic basemap, where higher intensity values correspond to greater SAR operational priority (see Figure 3).

Figure 3 provides an overview of the combined probabilistic surface, detected priority hotspot regions, and underlying incident reports within the interactive geospatial interface. Likewise, the full system architecture can be seen in Figure 4.

The development of this system architecture was carried out using R programming language [33]. Spatial preprocessing, surface generation, normalization, thresholding, and clustering are executed sequentially; when the hotspots metrics are computed these values are forwarded to the OpenAI model via LM Studio. Thus, this implementation runs in a local environment to ensure data privacy and operational reliability in post-earthquake scenarios. The truncated exponential and lognormal decay functions were implemented using the open source R package rgeoprofile, publicly available at [34]. Thus, the computational algorithm of the hybrid decision support framework is summarized in Algorithm 1 (Workflow of the proposed hybrid decision support framework).

It is important to note that the LLM used in this study does not directly classify the severity of incident reports from unstructured textual data. The LLM operates only after hotspot detection has been completed and receives structured quantitative indicators including hotspot rank, priority score, and operational radius. These numeric indicators guide the generation of recommendations, ensuring that the LLM works as a decision support interpreter rather than as a primary classifier of incident severity.

Algorithm 1 Combined Probabilistic SAR Prioritization
	Input: Incident dataset $\mathcal{E}={\left\{\left(x_i,t_i\right)\right\}}_{i=1}^N,$
	Output: Ranked hotspot regions Rk with LLM recommendations
1:	Load geolocated incident dataset
2:	Generate spatial grid over study area
3:	Compute TNE surface using truncated exponential decay
4:	Compute LN surface using lognormal decay
5:	Normalize both surfaces to [0, 1]
6:	Compute combined surface: PSAR = α·TNE + β·LN
7:	Determine threshold τ using upper quantile
8:	Create binary mask where PSAR ≥ τ
9:	Identify connected regions (clustering)
10:	for each region Rk do
11:	Compute maximum score Sk
12:	Compute area Ak
13:	Compute centroid ck
14:	end for
15:	Rank regions by Sk
16:	Select Top-K regions
17:	Generate structured LLM recommendations for each selected hotspot
18:	Return ranked regions and recommendations

3. Results

Hotspot Stability Under Alpha Variation

In order to assess the sensitivity of surface $P_{\mathrm{SAR}}\left(x\right)$, according to variations in the parameter α, the spatial model was tested with the following values, α = 0.5, α = 0.6, and α = 0.7, while keeping the quantile threshold $\tau =0.995$. As shown in Figure 5, increasing α value increases the influence of the truncated exponential function, resulting in greater spatial concentration around dense incident clusters. When α = 0.5, the surface exhibits a more diffuse spatial distribution with smoother gradients extending toward peripheral zones. At α = 0.6, the central cluster becomes more clearly defined, while α = 0.7 produces the most compact high-priority region, with a significant distinction between central and surrounding areas.

It should be noted that the top 3 hotspot regions remain geographically consistent across the three configurations. The centroids show minor positional changes, and the ranking order remains stable, suggesting robustness of the prioritization mechanism under moderate parameter variation.

From

Table 1. Summary of hotspot locations and emergency response recommendations generated by the LLM for the top 3 hotspots under α = 0.5, 0.6, and 0.7, including hotspot score, operational radius in meters, and full recommendation text.

α	Hotspot (Score)	Radius (Meters)	LLM Recommendation
0.5	HS1 (0.798)	261	PRIORITY: HIGH; TEAMS: 3; ACTIONS: deploy rapid response teams, conduct ground search with dogs and manual tools, use drones; RISKS: secondary collapse, aftershocks, unstable debris; TIME: HOURS
0.5	HS2 (0.772)	133	PRIORITY: HIGH; TEAMS: 4; ACTIONS: perimeter deployment, canine and thermal search, drone debris mapping, secure access routes; RISKS: aftershock destabilization, wall collapse, low visibility; TIME: HOURS
0.5	HS3 (0.763)	100	PRIORITY: HIGH; TEAMS: 2; ACTIONS: deploy units, establish triage within 100 m; RISKS: aftershocks, debris and traffic congestion; TIME: HOURS
0.6	HS1 (0.771)	157	PRIORITY: HIGH; TEAMS: 4; ACTIONS: rapid response, structural stability assessment, establish triage station; RISKS: further collapse, hazardous debris; TIME: IMMEDIATE
0.6	HS2 (0.762)	128	PRIORITY: HIGH; TEAMS: 3; ACTIONS: deploy ground teams, secure perimeter, establish communication hub; RISKS: secondary collapse, structural instability; TIME: HOURS
0.6	HS3 (0.754)	105	PRIORITY: HIGH; TEAMS: 3; ACTIONS: deploy units, triage perimeter, secure against aftershocks; RISKS: immediate aftershock damage, structural collapse; TIME: IMMEDIATE
0.7	HS1 (0.810)	274	PRIORITY: MEDIUM; TEAMS: 2; ACTIONS: rapid assessment team, systematic grid search; RISKS: aftershocks, instability; TIME: HOURS
0.7	HS2 (0.797)	100	PRIORITY: HIGH; TEAMS: 3; ACTIONS: rapid assessment, secure area, ground penetrating radar search; RISKS: aftershocks, gas leaks; TIME: IMMEDIATE
0.7	HS3 (0.794)	100	PRIORITY: HIGH; TEAMS: 4; ACTIONS: rapid search, establish medical stations, coordinate authorities; RISKS: structural collapse, hazardous material release; TIME: HOURS

, it can be seen that the LLM outputs shows strong consistency across all tested values of α. In every case, the model followed the prompt template by specifying priority level, number of teams, operational actions, identified risks, and response time.

The priority level was mainly categorized as HIGH across all configurations, with a single MEDIUM classification observed under α = 0.7 for the hotspot with the largest operational radius. Team allocation ranged between 2 and 4 units, without fluctuations with radius or score, suggesting that the LLM recommendations were conditioned not only to spatial extent but also on priority ranking. The operational actions showed standard urban SAR measures, including perimeter control, ground search deployment, structural stability assessment, triage establishment, and monitoring of aftershocks, indicating semantic coherence with the earthquake response scenario.

A comparative analysis across α values reveals that parameter variation affected both temporal urgency and tactical emphasis. For α = 0.5, all hotspots were classified with a HOURS response time, consistent with a more diffuse spatial surface. Under α = 0.6, two hotspots were labeled IMMEDIATE, indicating greater urgency when the truncated exponential component was more influential. For α = 0.7, urgency classifications were mixed, with one IMMEDIATE and two HOURS designations, while the highest score hotspot received a MEDIUM priority label. Although no linear relationship was observed between operational radius and team allocation, differences in hotspot size and score were shown in distinct tactical orientations. Larger areas tended to generate assessment oriented or systematic search strategies, whereas high-intensity zones tended to generate rapid response actions.

It is important to note that the structured correspondence between hotspot metrics and recommended actions suggests that the combined probabilistic surface influences the LLM outputs rather than producing generic advisory responses.

All the trials were carried out using a workstation with an Intel Core i7 9750H CPU at 2.60 GHz, 64 GB of RAM, and a 64 bit Windows 11 operating system. The full execution of the SAR prioritization pipeline, including spatial modeling, hotspot detection, ranking, and LLM-based recommendation generation, required approximately 13 to 15 min per run using the 40 incident dataset described in Appendix A. The truncated negative exponential model required on average 6 min to compute the spatial surface, while the lognormal decay model required approximately 7 min. It should be noted that the LLM-assisted decision support layer added some latency due to the calls to the local server. Each high-priority hotspot requiered approximately 1 to 2 min to generate emergency response recommendations. In summary, these results show that despite this processing time, our proposed framework is still practical for a post-eartquake scenario.

Even though the full pipeline execution required approximately 13–15 min, the proposed framework is intended to support early-stage post-earthquake situation rather than real-time response. In addition, it should be noted that this computational latency is due to the use of a local model running under the hardware constraints of these trials but the computational cost could be reduced by using optimized hardware or lighter LLM models.

4. Discussion

The results of this study show that the combined probabilistic spatial model is able to identify and generate operational recommendations for high-priority hotspots in complex urban environments such as Mexico City (RQ1). Moreover, the stability of hotspot detection across different values of the weighting parameter α indicates that the combined probabilistic spatial model provides a suitable mechanism for identifying high-priority SAR locations. On the other hand, the integration of a local LLM shows how quantitative hotspot indicators can be interpreted into structured operational recommendations, thereby enhancing the practical usability of spatial prioritization without compromising privacy or system transparency (RQ2).

While the current framework prioritizes spatial indicators from geolocated incident reports, real-world disaster response is strongly influenced by human behavior, community resilience, and social vulnerability [35]. In future implementations, this proposed framework could include different information sources such as survivor reports, emergency calls, and social media posts. Such data streams may provide insights into urgency, vulnerability, and community response patterns.

In real environments, the recommendations generated by the proposed framework should be interpreted with caution. The final decisions must remain under the supervision of professionals. A human verification layer is therefore needed to assess safety considerations. In this context, SAR coordination teams would confirm the suggested actions before any action, ensuring that algorithmic output remain aligned with operational protocols.

On the other hand, it is important to note that the interpretation of spatial prioritization results must consider human factors and risk perception as well [36]. Previous studies in seismic risk research have shown that community vulnerability, collective memory of past disasters, and public perception can influence how populations respond during emergency situations [37,38]. These psychosocial factors may shape evacuation behavior, and information dissemination. Although the proposed framework focuses on spatial indicators, integrating insights from human factor research may help contextualize hotspot prioritization within real environments.

5. Conclusions

This study set out to develop and assess a decision support framework integrating combined probabilistic spatial modeling with an LLM for post-earthquake search and rescue operations. The proposed approach transformed discrete incident reports into continuous operational priority fields. The sensitivity analysis demonstrated that hotspot identification remains stable under moderate variations of the weighting parameter α.

Future research may extend the proposed framework in several directions. Firstly, the integration of real-time information streams, such as social media posts, which may enhance the suitability of SAR prioritization during disaster events. Secondly, future work may explore systematic validation of recommendations generated by the local LLM through collaboration with SAR professionals and emergency management experts.