Abstract
Background
Social media plays an important role in disseminating information and identifying real-time needs during disasters. Its importance was shown on February 6, 2023, when Türkiye was struck by two destructive earthquakes. This resulted in tens of thousands of fatalities and economic devastation. In the early aftermath, social media was utilized as an important instrument for public coordination, information dissemination and social mobilization.
Methods
The study analyzed 697,313 tweets shared between February 6 and 14, 2023. Data were acquired via the X API v2 and processed to assess the engagement metrics of digitally expressed disaster needs using Python (version 3.12.7). Named Entity Recognition approach was applied to a combined rescue-request corpus to extract and geocode spatial data for heatmap visualization.
Results
Most frequent hashtags, types of requests, and patterns of interaction were investigated to examine how individuals communicated and expressed their needs during the earthquake. Findings showed that X contained time-sensitive, publicly generated signals related to rescue appeals, shelter needs, hygiene concerns and humanitarian coordination. During the first 12 hours of the earthquake, communication was primarily characterized by urgent requests for emergency rescue and heavy machinery. By the second day, shelter-related posts became more prominent with blanket and tent related requests. After the fourth day, there was a consistent rise in the posts about hygiene items and mobile toilets. The requests reached a peak between days five and seven.
Conclusion
The study highlights the potential value of social media data as a supplementary source for understanding publicly expressed needs during disaster response. X posts reflected a temporary shift from urgent rescue appeals to shelter, hygiene, and recovery-related concerns. However, because this study was descriptive and did not assess post accuracy, operational impact, or predictive performance, the findings should be interpreted as digital indicators of public concern rather than direct evidence of response effectiveness.
Keywords: Kahramanmaraş earthquakes, Social media and disaster response, Digital health communication, Crisis communication, Citizen generated data
Introduction
On February 6, 2023, two devastating earthquakes (magnitudes 7.8 and 7.6) struck southeast Türkiye, resulting in over 53,000 fatalities, the collapse of 9,432 buildings and the displacement of more than 2 million people [1]. The cost of the damage is expected to reach $100–155 billion with uncertain long-term socio-economic effects from disaster-induced domestic migration [2]. During the crisis, unique features of social media such as accessibility, mobility, and immediacy caused an increase in social media activity with numerous posts containing firsthand information. This allowed individuals to act as an information node, transforming an event into a globally visible incident almost instantly [3, 4]. As previous studies have shown, those affected used these platforms to request immediate help, supply delivery, debris removal, or search and rescue operations [5–7].
Among social media platforms, X is a practical tool for crisis response due to its real-time nature [8, 9]. With over 415.3 million active users and 400 million daily tweets, X serves as a crowdsourcing platform for information dissemination during disasters. It enhances situational awareness for decision makers [10, 11]. An expanding body of literature examines X data to explain public behaviour during or after disasters [12, 13]. These studies describe crisis-related information propagation and analyze the varied range of information regarding the features of the event. It also allows people to spread information instantly and coordinate resources by improving situational awareness. X was utilized by government and non-governmental organizations (NGOs) as a real-time monitoring system in events such as the Indonesian tsunamis, California wildfires, and Australian floods [14, 15].
By enabling users to categorize content and engage in rapid information exchange beyond their networks, hashtags have become important features of digital interaction [16]. While this functionality extends audience reach and facilitates real-time communication, it introduces challenges such as unverified information, rumors, and imprecise data [17, 18]. Despite this, X is the most investigated platform in social media analysis due to its data volume, speed, and simplicity [19]. It is conceptualized as a “social sensor network”, where the collective input of distributed human sensors provides a real-time, crowdsourced data stream essential for evaluating and directing field operations.
Effective public health communication in such contexts requires immediate and focused communication anchored in the “3 Rs” framework: “Review, Recognize and Respond”. This principle ensures that health communication is data-driven, goal-oriented, and meaning-driven [20]. The examination of social media data provides a view into the psychological states and urgent public concerns. However, dissemination of unverified information or emotional content such as the significant shock and prayer tweets following the Kahramanmaraş event, poses risks and can escalate collective anxiety [21, 22]. Consequently, disaster policies should integrate mechanisms to monitor psychological sensitivity and combat misinformation on social media platforms [23].
Disaster preparedness should be expanded to encompass digital resilience. Unlike traditional models that focus on physical infrastructure and logistics, digital preparedness emphasizes the reliability of communication systems [24, 25]. This framework requires clear strategies for pre-planned messaging and the management of misinformation. During the pre-crisis stage, social media is instrumental in identifying early warning signs and building public trust. However, during emergencies, the focus shifts: alerts must be brief and actionable. Effective crisis communication should consider the significant cognitive pressure imposed on the public by prioritising practical, verified information rather than secondary details [26, 27]. The contribution of this study is therefore context-specific and empirical. First, it describes the temporal sequence of publicly expressed needs during the 2023 Kahramanmaraş earthquakes, including the shift from urgent rescue appeals to shelter, hygiene, sanitation, and recovery-related concerns. Second, it compares tweet volume and engagement patterns to distinguish frequently mentioned needs from highly amplified needs. Third, it applies and manually validates a Turkish language usable location detection workflow for non-geotagged rescue-related posts.
Previous studies in crisis informatics have shown that social media facilitates communication and enhances situational awareness during disasters. However, the evolution of public requirements within Turkish-language social media during the 2023 Kahramanmaraş earthquakes is not well understood. In particular, the transition from urgent rescue appeals to longer-term concerns regarding shelter, hygiene, and recovery has not been sufficiently documented. Additionally, the process of extracting location data from non-geotagged Turkish posts remains a challenge that requires empirical attention. This study investigates X usage during Kahramanmaraş earthquakes by analyzing message volume, content, and engagement patterns. It also traces the timing and geographic distribution of disaster-related posts to understand how affected communities expressed their needs, rather than to claim novelty for the general concept of social media situational awareness or to measure operational effectiveness or predictive accuracy of social media monitoring.
Material and method
This study was designed as a descriptive analysis of open public X posts. It did not assess the accuracy of individual posts, compare social media signals with official field records, or perform predictive modeling. The analytical framework focused on identifying patterns of digitally expressed needs and analyzing observed engagement within the sample to better understand public communication during the disaster.
Tweets were collected between February 6 and 14, 2023, through X API v2 using Python’s Tweepy library. Data retrieval was based on predefined earthquake-related hashtags and phrases listed below. A post was eligible for inclusion if it was publicly available, posted during the study period, and contained at least one of the predefined search terms. The search query was constructed using boolean logic to capture any public tweet containing at least one of the predefined list. The data collection resulted in 721,783 tweets. To isolate original posts, retweets and replies were excluded directly during the data retrieval process using the API filtering. Quote tweets, which include user commentary alongside shared content, were retained as original posts. Retweets were excluded as separate observations because they reproduced the content of the original posts; however, the retweet and like counts of each retained original tweet were preserved as engagement variables. Because some search terms were province names (e.g., #Hatay, #Gaziantep), initial retrieval also contained content unrelated to the earthquake response, such as advertisements, city-promotion campaigns, photographs, and other noisy X streams. To address this issue, the dataset was processed using the Python Pandas library. We applied a custom exclusion dictionary to identify and remove irrelevant posts based on text content. Duplicate records were removed based on the tweet’s unique ID number. Consequently, a total of 24,470 non-usable records were identified and removed. Although the total number of excluded records was retained, the specific exclusion counts for each category were not preserved during preprocessing. The final sample consisted of 697,313 publicly available and usable tweets. For each original tweet, the dataset included tweet’s unique ID, user ID, message, post date, number of retweets, and number of likes. Location data were not available in the X dataset.
To identify the messages about the earthquake, the following hashtags/phrases were used: #deprem (earthquake), #depremoldu (earthquake occurred), #Gaziantep, #Hatay, #AFAD (Disaster and Emergency Management Presidency), #Malatya, #Adıyaman, #yardım (help), #enkaz (rubble), #kızılay (Red Crescent), #enkazaltındayım (I’m under the rubble), #iskenderun, #Kahramanmaraş, #Ahbap (The Ahbap Association), #seferberlik (campaign), #AKUT (Search and Rescue Association), #hatayyardimbekliyor (hatay is waiting for help), #ohal (State of Emergency Rule), çok acil (very urgent), #depremsondakikadeprem (earthquake breaking news), teyitli bilgi (confirmed information), and #acildeprem (earthquake emergency).
Trending topics data was provided from X API’s GET trends/place function that returns the top 50 trending topics at the country level. The trend lists were created based on tweet volumes which are array data of trend objects in the X API’s function. The most up-to-date information available is returned on request. The X API v2 also gave access to volume streams, which can identify and track trends in a given time period.
We used publicly available rescue requests to create a heatmap visualizing the concentration of collapsed buildings. Data were retrieved from the Deprem platform (deprem.io) and the posts from X. Tweets were selected to reflect emergency situations such as “I’m under rubble”, “collapsed building”, “help”, or “address”. These combined data sources were merged into a single corpus. The initial corpus was raw and unstructured, filled with stop words, URLs, and symbols. We used the Python NLTK library to achieve a clean and consistent format. This involved converting all text to lowercase, removing punctuation and non-textual elements, and applying tokenization and lemmatization. A hybrid approach was utilized to identify addresses, combining a Named Entity Recognition model with custom rule-based patterns. Multiple posts could originate from a single collapse. For this reason, we implemented a deduplication strategy. Reports that geocoded to within a 50-meter radius were clustered and treated as a single incident to prevent over-representing any one location. The final list of unique addresses was then converted into longitude and latitude coordinates using the Google Maps Geocoding API. Finally, these coordinates were visualized using the ggmap library in R. The resulting heatmaps highlight the main hotspots within the earthquake zones, where the density of rescue requests was highest.
Manual validation of usable location detection
To examine the reliability of the location extraction process, two human reviewers independently evaluated a validation sample of 673 posts, and disagreements were resolved by consensus. The validation sample was not stratified according to province, date, or time period. Reviewers evaluated whether each post contained usable location data, defined as spatial information more specific than a general provincial reference that permitted geocoding. Posts that contained unclear, general, or unverifiable location references were coded as missing location data. Algorithm outputs were compared with the human-reviewed classifications, and precision, recall, F1-score, accuracy, and specificity were calculated.
Results
The study reports three quantitative sources: (1) counts, classifications, and engagement metrics from the sample of 697,313 original tweets; (2) platform-level trend indicators obtained from X API trend functions; and (3) the combined rescue-request corpus used for spatial mapping. These measures were employed for different analytic purposes and were not interpreted interchangeably.
Table 1 displays the ten most popular hashtags and phrases in the first 12 hours after the earthquake (The earthquake occurred at 04:17 TRT, February 6, 2023). A sharp rise in tweets was observed immediately after the earthquake. Following the earthquake, there was an increase in the frequency of tweets containing phrases such as “an earthquake occurred”, “get well soon”, “help”, “collapsed building”, “address”, “AFAD”, and “emergency”.
Table 1.
Trending hashtags and phrases in the first 12 hours (platform-level volume)
| Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |
#depremoldu (earthquake occurred) |
147.5 |
#depremoldu (earthquake occurred) |
1009.6 | #deprem (earthquake) | 3754.5 | #deprem (earthquake) | 5222.1 | |||
| 2 | #geçmişolsun (get well soon) | 56.8 | #deprem (earthquake) | 564.3 | göçük (dent) | 577.5 | göçük (dent) | 704.4 | |||
| 3 | #Kahramanmaraş | 54.1 | #yardımedin (help) | 489.8 | #AFAD* | 450.2 |
#enkazaltındayım (I’m under the rubble) |
176.3 | |||
| 4 | #Diyarbakır | 43.7 | göçük (dent) | 334.6 |
#enkazaltındayım (I’m under the rubble) |
88.5 | #acil (emergency) | 165.7 | |||
| 5 | #Elazığ | 27.2 | #Hatay | 287.4 | #Malatya | 83.0 | Allah’ım (oh my god) | 103.9 | |||
| 6 | #TümTürkiye (All Türkiye) | 26 | #AFAD* | 261.1 | yıkılan bina (collapsed building) | 79.5 | adres (address) | 98.3 | |||
| 7 | Allah’ım (oh my god) | 20.4 | Merkez üssü (epicenter) | 96.3 | #earthquake | 71.7 | #earthquake | 91.4 | |||
| 8 | #Malatya | 16.2 | #acil (emergency) | 79.4 | adres (address) | 69.3 | #Şanlıurfa | 60.5 | |||
| 9 | Türkiyem (my Türkiye) | < 10 K | #Diyarbakır | 68.7 | #kandilli** | 46.3 | #Turkey | 58.0 | |||
| 10 | #iyiyiz (we are safe) | < 10 K | #Malatya | 47.0 | #Şanlıurfa | 42.8 | #kandilli** | 56.6 | |||
| # | 2023.02.06 (8:00 a.m. − 9:00 a.m.) | 2023.02.06 (9:00 a.m. − 10:00 a.m.) | 2023.02.06 (10:00 a.m. − 11:00 a.m.) | 2023.02.06 (11:00 a.m. − 12:00 a.m.) | |||||||
| 1 | #deprem (earthquake) | 2170.5 | #deprem (earthquake) | 2239.8 | #deprem (earthquake) | 3938.8 | #deprem (earthquake) | 4797.6 | |||
| 2 |
#enkazaltındayım (I’m under the rubble) |
229.1 |
#enkazaltındayım (I’m under the rubble) |
269.3 |
#enkazaltındayım (I’m under the rubble) |
325.5 | sokakno (street number) | 650.3 | |||
| 3 | #acil (emergency) | 213.3 | #acil (emergency) | 250.8 | #acil (emergency) | 280.2 |
#enkazaltındayım (I’m under the rubble) |
402.9 | |||
| 4 | #earthquake | 118.2 | arkadaşımın (my friend’s) | 220.3 | arkadaşımın (my friend’s) | 245.6 | #acil (emergency) | 302.9 | |||
| 5 | #Turkey | 86.1 | #earthquake | 155 | #earthquake | 182.1 |
arkadaşımın (my friend’s) |
287.7 | |||
| 6 | #Şanlıurfa | 77.5 | Allah’ım (oh my god) | 137.6 | Allah’ım (oh my god) | 162.9 | Allah’ım (oh my god) | 266.3 | |||
| 7 | saat 04.17 (04.17 a.m.) | 60.1 | #Şanlıurfa | 92.2 | saat 04.17 (04.17 a.m.) | 91.3 | Odabaşı | 220.9 | |||
| 8 | #kandilli** | 55.0 | saat 04.17 (04.17 a.m.) | 81.5 | #kandilli** | 67.1 | #Malatya | 195.8 | |||
| 9 | Allah’ım (oh my god) | 50.4 |
#sondakika (breaking news) |
68.2 | #İskenderun | 63.2 | #Turkey | 161.0 | |||
| 10 | #HAARP*** | 40.4 | #kandilli** | 62.9 | #Nurdağı | 57.8 | #Şanlıurfa | 115.6 | |||
| # | 2023.02.06 (12:00 a.m. − 1:00 p.m.) | 2023.02.06 (1:00 p.m. − 2:00 p.m.) | 2023.02.06 (2:00 p.m. − 3:00 p.m.) | 2023.02.06 (3:00 p.m. − 4:00 p.m.) | |||||||
| Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | ||||
| 1 | #deprem (earthquake) | 5721.0 | #deprem (earthquake) | 2724.9 | #deprem (earthquake) | 2699.2 | #deprem (earthquake) | 2656.2 | |||
| 2 | sokakno (street number) | 761.8 | bina (building) | 783.1 | bina (building) | 839.4 | bina (building) | 884.2 | |||
| 3 | arkadaşımın (my friend’s) | 336.8 | sokakno (street number) | 749.8 | sokakno (street number) | 721.7 | sokakno (street number) | 821.5 | |||
| 4 | #acil (emergency) | 310.2 | arkadaşımın (my friend’s) | 372.1 | arkadaşımın (my friend’s) | 419.9 |
#enkazaltındayım (I’m under the rubble) |
528.5 | |||
| 5 | Odabaşı | 256.1 | #acil (emergency) | 337.5 | #acil (emergency) | 366.1 | arkadaşımın (my friend’s) | 465.0 | |||
| 6 | #Turkey | 217.3 | Odabaşı | 287.1 | #Turkey | 328.2 | #Turkey | 391.3 | |||
| 7 | #Malatya | 215.1 | #Turkey | 265.5 | #earthquake | 315.1 | Odabaşı | 342.2 | |||
| 8 | #Şanlıurfa | 127.5 | Onikişubat | 210.1 | Odabaşı | 309.9 | #kızılay (redcrescent) | 179.7 | |||
| 9 | #kızılay (redcrescent) | 111.2 | #enkaz (rubble) | 140.9 | #Akevler | 260.8 | #Şanlıurfa | 151.1 | |||
| 10 | #enkaz (rubble) | 104.3 | #İskenderun | 112.5 | #kızılay (redcrescent) | 156.3 | #afet (disaster) | 133.6 | |||
English equivalents of the words are shown in parentheses. Terms without parentheses are the hashtags in English and place names such as province, district, street etc. Values are shown in thousands and represent platform-level trend volume indicators returned by X API trend functions for the listed terms during each interval. They do not correspond to counts within the analyzed sample of 697,313 original tweets
*Disaster and Emergency Management Presidency of Türkiye
**Kandilli Observatory and Earthquake Research Institute
*** High-frequency Active Auroral Research Program
Two hundred twenty nine point one thousand tweets containing “I am under rubble” began to emerge and rose to the top of the list of most frequent tweets, between 8 and 9 am. Requests for foreign aid started to increase from 8 am onwards. Between 9 and 10 am, users started to send tweets requesting help for their friends trapped under the rubble, using the phrases starting with “my friend’s address/name….” and province, district, street and building names. Additionally, the tweets containing prayers, moral support, and motivation that began with “Oh my God” were among the top 10 hashtags. These tweets expressed shock and prayers regarding the severity of the earthquakes. After 11 am, tweets requesting help, including building and street addresses, began to feature in the top 10 tweets. According to platform-level volume streams, the 12 − 1 pm interval had the highest platform-level trend volume for earthquake-related terms, totaling approximately 7 million across the listed terms globally. The second earthquake with a magnitude of 7.6 that occurred in Kahramanmaraş at 01:24 pm was reflected in the tweets posted between 1 and 2 pm. The tweets starting with phrases such as “my friend’s address/name…” showed an increasing trend. As aftershocks occured, tweets with phrases like “aftershock” rose to the top. Place names such as Odabaşı (in Hatay province), Nurdağı (in Gaziantep province), and Onikişubat (in Kahramanmaraş province) appeared frequently among the top terms in several hourly intervals. These findings indicate high mention frequency within the sampled X posts and should not be interpreted as independent evidence of damage severity, rescue demand, or structural collapse in these locations. At 2:00–3:00 pm, hashtags such as “let’s donate blood” did not rank high, whereas “spread, please” took the lead. In addition, the tweets tagged with the names of cities affected by the earthquake were observed in the top trending hashtags within every hour.
The most popular hashtags in the hours following the first 12 hours after the earthquake were given in Table 2. The tweets reporting that people’s relatives or friends were under the rubble (I’m under rubble, still under rubble, my friend’s address…) had the highest number of retweets at almost every time interval until 4 am. From 4 to 5 pm, tweets related to “government and people working together”, “Turkish Red Crescent”, and “blankets” became prominent. While Odabaşı tweets continued to appear as top trending tweets at many hours, tweets about “Nurdağı” district (one of the epicenters of the earthquakes) took the lead at 7–8 pm. The tweets calling for “the cancellation of transactions in the stock market” and “the stock market’s closure until February 20th” gained significant attention and ranked high in popularity. The hashtag “Haluk Levent”, the founder of Ahbap foundation and famous rock musician, was among the top 10 between 8 and 9 pm. After this hour, the Turkish Red Crescent consistently ranked high in popularity.
Table 2.
Trending hashtags and phrases in the second 12 hours following the earthquake (Platform-level volume)
| # | 2023.02.06 (4:00 p.m. − 5:00 p.m.) | 2023.02.06 (5:00 p.m. − 6:00 p.m.) | 2023.02.06 (6:00 p.m. − 7:00 p.m.) | 2023.02.06 (7:00 p.m. − 8:00 p.m.) | |||||
|---|---|---|---|---|---|---|---|---|---|
| Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | ||
| 1 | #deprem (earthquake) | 2566.5 | #deprem (earthquake) | 2341.3 | #deprem (earthquake) | 5112.5 | #deprem (earthquake) | 2876.5 | |
| 2 | bina (building) | 925.6 | arkadaşımın (my friend’s) | 1041.6 | #Gaziantep | 743.3 | hala enkazda (still under rubble) | 1053.5 | |
| 3 |
#enkazaltındayım (I’m under rubble) |
570.2 | #Gaziantep | 731.6 |
#enkazaltındayım (I’m under rubble) |
629.8 | #Turkey | 686.1 | |
| 4 | arkadaşımın (my friend’s) | 497.9 |
#enkazaltındayım (I’m under rubble) |
607.8 | #Turkey | 608 | arkadaşımın (my friend’s) | 617 K | |
| 5 | #Turkey | 451.4 | #Turkey | 527.3 |
arkadaşımın (my friend’s) |
576.5 |
Lütfen yayalım (spread, please) |
528.4 | |
| 6 |
lütfenyayalım (spread, please) |
403.7 |
Lütfen yayalım (spread, please) |
442.4 | Lütfen yayalım (spread, please) | 487.3 | Odabaşı | 478.5 | |
| 7 | Odabaşı | 362.8 | Odabaşı | 397.0 | Odabaşı | 436.5 | #kızılay (redcrescent) | 317.5 | |
| 8 | #kızılay (redcrescent) | 203.8 | #kızılay (redcrescent) | 230.2 | #kızılay (redcrescent) | 279.3 | #battaniye (blanket) | 200.0 | |
| 9 | #afet (disaster) | 149.3 | #afet (disaster) | 164.5 | #afet (disaster) | 177.7 | #afet (disaster) | 190.3 | |
| 10 | #artçıdeprem (after shock) | 121.3 | #İskenderun | 158.9 | #battaniye (blanket) | 170.5 | #Nurdağı | 147.3 | |
| # | 2023.02.06 (8:00 p.m. − 9:00 p.m.) | 2023.02.06 (9:00 p.m. − 10:00 p.m.) | 2023.02.06 (10:00 p.m. − 11:00 p.m.) | 2023.02.06 (11:00 p.m.- 12:00 p.m.) | |||||
|---|---|---|---|---|---|---|---|---|---|
| 1 | #deprem (earthquake) | 1685.9 | #deprem (earthquake) | 4304.4 |
#enkazaltındayım (I’m under rubble) |
2236.5 |
#enkazaltındayım (I’m under rubble) |
2479.7 | |
| 2 |
hala enkazda (still under rubble) |
1220.6 |
#enkazaltındayım (I’m under rubble) |
2123.2 | #Hatay | 2054.1 | #Turkey | 1000.1 | |
| 3 | #Turkey | 764.6 | #Turkey | 841.8 | #Turkey | 910.4 | arkadaşımın (my friend’s) | 772.5 | |
| 4 | arkadaşımın (my friend’s) | 658.7 | arkadaşımın (my friend’s) | 701.7 | arkadaşımın (my friend’s) | 738.7 | #Nurdağı | 191.1 | |
| 5 | #earthquake | 511.7 | #Nurdağı | 170.4 | #kızılay (redcrescent) | 411.2 | #haluklevent (volunteer) | 109.9 | |
| 6 | #kızılay (redcrescent) | 350.1 | #nacigörür (geoscientist) | 121.1 | #nacigörür (geoscientist) | 139.1 | #seferberlik (campaign) | 87.8 | |
| 7 | #Nurdağı | 159.5 | #haluklevent (volunteer) | 70.8 | #Nurdağı | 175.5 | #oguzhanugur (youtuber) | 77.3 | |
| 8 | #nacigörür (geoscientist) | 103.1 | asker (soldier) | 61.3 | Yardım edin (help) | 98.4 | #celalsengor (geoscientist) | 76.1 | |
| 9 | #haluklevent (volunteer) | 55.6 | # AKUT* | 46.1 | #haluklevent (volunteer) | 94.2 | #AKUT* | 60.9 | |
| 10 | asker (soldier) | 51.1 | #ohal** | 45.7 | #ohal** | 63.9 | #sela (call to prayer) | 59.7 | |
| # | 2023.02.07 (12:00 a.m. − 1:00 a.m.) | 2023.02.07 (1:00 a.m. − 2:00 a.m.) | 2023.02.07 (2:00 a.m. − 3:00 a.m.) | 2023.02.07 (3:00 a.m.- 4:00 a.m.) | |||||
|---|---|---|---|---|---|---|---|---|---|
| Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | Hashtag/phrase | Number of Tweets | ||
| 1 |
#enkazaltındayım (I’m under rubble) |
2285.6 |
#hatayyardımbekliyor (Hatay is waiting for help) |
1464.8 | #Hatay | 1326.4 | #Hatay | 1495.7 | |
| 2 | #Hatay | 2208.3 | #Turkey | 1220.3 | #Turkey | 1265.1 | #Turkey | 1313.9 | |
| 3 | #Turkey | 1097.6 | arkadaşımın (my friend’s) | 840.5 |
#hatayyardımbekliyor (Hatay is waiting for help) |
850.2 |
kurtarma çalışmaları (rescue works) |
973.3 | |
| 4 | arkadaşımın (my friend’s) | 784.5 |
#enkazaltındayım (I’m under rubble) |
554.3 |
#enkazaltındayım (I’m under rubble) |
591.2 |
#hatayyardımbekliyor (Hatay is waiting for help) |
927.9 | |
| 5 | #kızılay (redcrescent) | 443.7 | #kızılay (redcrescent) | 520.6 | #kızılay (redcrescent) | 551.5 |
#enkazaltındayım (I’m under rubble) |
604.8 | |
| 6 | #Nurdağı | 200.3 | #İskenderun | 285.9 | #ohal** | 254.3 | #kızılay (redcrescent) | 484.4 | |
| 7 | #ohal** | 146.4 | #ohal** | 217.9 | #seferberlik (campaign) | 244.8 | #ohal** | 270.1 | |
| 8 | #haluklevent (volunteer) | 126.8 | #Nurdağı | 207.6 | #Nurdağı | 211.3 | #seferberlik (campaign) | 266.7 | |
| 9 | #oguzhanugur (youtuber) | 90.6 | #seferberlik (campaign) | 201.9 | #haluklevent (volunteer) | 143.4 | #Nurdağı | 213.5 | |
| 10 | #celalsengor (geoscientist) | 89.2 | #haluklevent (volunteer) | 139.9 | Ekinci | 140.3 | #haluklevent (volunteer) | 148.6 | |
English equivalents of the words are shown in parentheses. Terms without parentheses are hashtags in English and place names such as province, district, street etc. Values are shown in thousands and represent platform-level trend-volume indicators returned by X API trend functions for the listed terms during each interval. They do not correspond to counts within the analyzed sample of 697,313 original tweets
*Search and Rescue Association in Türkiye
** State of Emergency Rule
Table 3 shows the ten most popular daily trends in Türkiye from February 7 to 14, 2023. The most popular tweets on the second and third days following the earthquake were pleas for help such as “I am under rubble”, “Hatay and Adıyaman need rescue”, and “There is a voice”. During this period, the top 10 tweets included hashtags: “Red Crescent”, “Ahbap Association”, “AKUT”, and well-known public figures such as “Haluk Levent”, and “Oğuzhan Uğur”. These posts received high visibility in the sampled X discourse and reflected public attention to rescue appeals, humanitarian organizations, and well-known public figures. During the third and fourth days, tweets increasingly addressed necessities such as shelter/housing, drinking water, medication, cranes, and mobile toilets. The subject of housing earthquake victims in state-owned dorms and the shutdown of colleges was also raised, with some of the most popular tweets expressing a desire for schools to stay open and opposing distance learning. Finally, among the most popular tweets in the first week were those supporting the earthquake-affected region, such as “We are together,” “Pray for Turkey,” and “Türkiye One Heart”.
Table 3.
Most popular trends in Türkiye from February 7 to 14, 2023
| # | 2023.02.07 | 2023.02.08 | 2023.02.09 | 2023.02.10 | 2023.02.11 | 2023.02.12 | 2023.02.13 | 2023.02.14 |
|---|---|---|---|---|---|---|---|---|
| 1 | #Turkey | #seferberlik (campaign) |
#sesvar (voice from the rubble) |
#kızılay (redcrescent) |
#depremzede (victims) | #oguzhanugur (youtuber) |
#borsadaislemler iptaledilsin (cancel transactions on stock market) |
#borsadaislemler iptaledilsin (cancel transactions on stock market) |
| 2 | #hatayyardım bekliyor (Hatay is waiting for help) |
#sesvar (voice from the rubble) |
#acildeprem (emergency earthquake) |
#AKUT | #yağma (plunder) |
#okullarçılsın (open schools) |
Canımız Türkiye (Our dear Türkiye) |
Ped Ped Ped (sanitary pad) |
| 3 |
#enkazaltındayım (I’m under rubble) |
#sondakikadeprem (breakingearthquake) |
#ArmyAssemble ForTurkey |
#AhbapDernegi | #AhbapDernegi |
#borsadaislemler iptaledilsin (cancel transactions on stock market) |
#BeraberizTürkiye (Together Türkiye) |
#universiteler kapatılmasın (don’t close the universities) |
| 4 |
#kızılay (redcrescent) |
#AdıyamanYardım Bekliyor (Adıyaman is waiting for help) |
#Hatayiskenderun | #haluklevent (volunteer) |
#uzaktanegitim (distance- education) |
BabalaTv (Youtube Channel) |
#universiteler kapatılmasın (don’t close the universities) |
#SeçimlerErtelenemez (elections cannot be postponed) |
| 5 | #İskenderun | AcilYardım (emergencyhelp) |
#TwitteriAçın (Unban Twitter) |
#oguzhanugur (youtuber) |
#MobilTuvalet (mobile toilet) |
#Kahramanmaraş |
#EğitimDevamEtmeli (Education must continue) |
#EğitimDevamEtmeli (Education must continue) |
| 6 | #ohal (state of emergency rule) | Odabaşı | #oguzhanugur (youtuber) |
Arzu apartmanı (Arzu apartment) |
#polis (police) |
#baykar (private business) |
#Borsa20Subata KadarKapatılsın (Stock must be closed until February 20) |
Devletimiz Yanımızda (our state with us) |
| 7 | #seferberlik (campaign) |
Habertürk (TV Channel) |
#AhbapDernegi |
#Nasuhmahruki (volunteer) |
# SuGötürün (take water to region) |
#Konut Seferberliği (Housing campaign) |
PED PED PED (hijyenic pad) |
BabalaTv (Youtube Channel) |
| 8 | #Nurdağı |
# SuGötürün (take water to region) |
#KefenLazım (Need for shrouds) |
#AFAD | #imar (reconstruction) | #imar (reconstruction) |
Selçuk Bayraktar (businessman) |
#uzaktanegitim istemiyorum (no distance education) |
| 9 | #haluklevent (volunteer) |
#acilvinc (emergency crane) |
TahaDuymaz YardımBekliyor (Taha Duymaz is waiting for help) |
#depremzede (victims) |
#TürkiyeTek Yürek (Türkiye One Heart) |
#soruşturma (investigation) |
Babala TV (Youtube Channel) |
#kiralardüşsün (let the rents fall) |
| 10 | #oguzhanugur (youtuber) |
eczaneler (pharmacies) |
#Gaziantep deprem (Gaziantep earthquake) |
#suriyeliler (syrians) |
Deniz Baykal (politician) |
#kykçözümdeğil (credit and dormitories institution is not a solution) |
#TBMM (Grand National Assembly of Türkiye) |
#Borsa20SubataKadar Kapatılsın (Stock must be closed until February 20) |
Hashtags and phrases were extracted from the sample data (n = 697,313). It presents daily platform-level trend lists obtained from X API trend functions during February 7–14, 2023
Table 4 summarizes engagement metrics for the classified subset of tweets assigned to selected high-frequency hashtag/phrase categories. Of the full sample, 636,055 tweets were assigned to one of the three categories, while 61,258 tweets were not categorized because they were outside the selected top terms or contained multiple category-defining hashtags. Therefore, category-level results refer to the classified subset only. In this subset, propaganda/campaign hashtags had lower average retweet and like values than humanitarian-support and some request/need hashtags. The hashtag “#85MilyonTekYürek” (85 million one heart) had the highest number of tweets within the propaganda/campaign category, while “#DevletMilletElele” (state and nation together) had the highest average retweet and like counts within that category. Humanitarian- agency and public-figure hashtags also showed high engagement. Within the requests and needs category, blanket-related posts were more frequent, whereas tent-related posts had higher average retweets.
Table 4.
Most frequent hashtags and phrases by categories (n = 636,055)
| Propaganda & campaign (n = 143,242) | Humanitarian support, services (n = 417,627) | Requests and needs (n = 75,186) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Hashtag/phrase | n |
|
|
Hashtag/phrase | n |
|
|
Hashtag/phrase | n |
|
|
|
#85MilyonTekYürek (85 million one heart) |
20,509 | 1.01 | 0 |
#kızılay (redcrescent) |
174,328 | 9.89 | 1.42 | battaniye (blanket) | 43,582 | 12.10 | 2.26 |
|
BirlikteGüçlüyüz (We together strong) |
17,433 | 8.81 | 0 | #AFAD | 63,392 | 442.59 | 2.87 |
#borsadaişlemleriptaledilsin (cancel transactions on the stock market) |
12,913 | 5.68 | 1.28 |
| #PrayForTurkey | 15,496 | 7.76 | 3.05 | #AhbapDernegi | 34,866 | 21.77 | 1.22 | #Kepçevinç (grab and crane) | 6457 | 11.84 | 2.42 |
|
#TürkiyeTekYürek (Türkiye One Heart) |
18,350 | 21.36 | 3.02 | #haluklevent (volunteer) | 24,904 | 7.66 | 1.07 | #DepremBölgelerineKefenLazım (need of shrouds for earthquake region) | 2893 | 9.03 | 0.80 |
|
#BizBeraberiz (We are together) |
17,008 | 0.28 | 3.01 | #oguzhanugur (youtuber) | 19,923 | 230.81 | 54.70 |
#DepremBölgelerineÇadırYardımı (tent aid to earthquake region) |
2257 | 180.01 | 1.27 |
|
#HaydiKanVer (let’s donate blood) |
13,946 | 3.23 | 1.67 | #AKUT | 16,217 | 16.73 | 0.53 |
#DepremBölgelerine MobilTuvalet (mobile toilet for earthquake zones) |
1506 | 15.42 | 0.81 |
|
#DevletMilletElele (state and nation together) |
14,527 | 22.51 | 10.05 | #nacigörür (geoscientist) | 46,488 | 2.40 | 0.75 | #SuGötürün (take water to region) | 1484 | 2.63 | 0.19 |
|
#TürkiyeİçinİyilikVakti (favor time for Türkiye) |
10,565 | 5.39 | 1.22 | #celalşengör (geoscientist) | 24,045 | 31.55 | 10.52 |
#PişikKremiBebekBezi (baby diapers and rash cream) |
1462 | 18.90 | 0.76 |
|
Devletimiz yanımızda (Our state is with us) |
8504 | 8.62 | 0.01 |
#nasuhmahruki (volunteer) |
7748 | 123.92 | 0.21 |
#KızlarİçinPedGönderin (send sanitary pad for women) |
1362 | 25.71 | 0.44 |
|
Devletimin yanındayım (I am with my state) |
6904 | 9.56 | 5.11 |
#AnkaraBB (Ankara Metropolitan Municipality) |
5716 | 0.04 | 0.67 |
#DepremBölgelerineMaske (need of mask for earthquake region) |
1270 | 2.58 | 0.13 |
|
Average (per tweet) |
14324.21 | 8.79 | 2.59 |
Average (per tweet) |
41762.77 | 89.62 | 4.52 |
Average (per tweet) |
7518.63 | 15.99 | 1.85 |
Table 4 includes 636,055 tweets corresponding to 91.2% of the full sample. Tweets outside the selected top hashtags/phrases or containing multiple category defining hashtags were not assigned to a single category to avoid double counting (n = 61,258; 8.8%). Therefore, category level counts and engagement metrics refer to the classified subset only and may underrepresent less frequent or multitopic posts
n indicates the total number of tweets that included the hashtag or phrase in the sample
symbol indicates average number of retweets and
symbol is for average number of likes per tweet
Figure 1 shows the percentage of sampled posts in each time interval that contained selected request/need categories, from the first hour after the earthquake through Day 9. Requests for heavy equipment, such as grab and crane, were most visible during the first 24 hours, when search-and-rescue-related discourse was prominent. By contrast, tweets requesting blankets and tents peaked later, especially on the second and third days, when large numbers of displaced people were trying to secure temporary shelter. After Day 4, the signal shifted toward basic living conditions. Posts asking for mobile toilets and menstrual or hygiene supplies became more prominent between Days 5 and 7, suggesting an increased online attention to sanitation-and-dignity-related concerns.
Fig. 1.
For each time point, all sampled tweets posted during that interval were screened for the selected request/need categories. Percentages were calculated as the number of tweets in a given category during that time interval divided by the total number of sampled tweets posted during the same interval, multiplied by 100. Time points with no tweets in the selected categories were coded as 0%. Because only selected request/need categories are shown, the percentages do not represent the full distribution of all earthquake-related tweets. Hashtags were originally in Turkish and written as compound words (e.g., “mobiltuvalet,” “kizlaricinpedgonderin”). For clarity, they were translated into English and formatted as readable phrases (e.g., “mobile toilet for earthquake region,” “send sanitary pad for women”). Distinct grayscale line styles are used for print readability
Analysis of the combined rescue-request corpus derived from X posts and publicly available deprem.io records produced two visualizations of digitally reported aid-request concentrations: Fig. 2 illustrates the hotspots of aid requests across the four major affected provinces (Kahramanmaraş, Osmaniye, Hatay, and Malatya), while Fig. 3 shows the hotspots of aid requests from the earthquake zone. Within the analyzed dataset, Hatay was the most frequently mentioned province in posts requesting help. Digitally reported rescue requests were concentrated mainly in city and district centers. This pattern may reflect population density, infrastructure damage, internet access, and differences in social media use, and should not be interpreted as evidence that rural areas had lower levels of need.
Fig. 2.
Hotspots of aid requests from the earthquake zones: A Kahramanmaras, B Osmaniye, C Hatay, and D Malatya provinces, in order from top right to bottom left
Fig. 3.
Hotspots of aid requests from the whole earthquake zone. Heat map data were derived from the tweets and publicly available the Deprem platform’s (deprem.io) messages requesting help, which included the addresses of collapsed buildings and rubble sites. Longitude/latitude coordinates of the addresses were used for visualizing spatial data. The spatial information from a static Google Maps layer was combined with the ggmap library in R. The maps indicate the earthquake zones containing the primary the hotspots of activity. Hotspots on the maps correspond to locations with a high density of digitally reported rescue requests mentioning collapsed buildings or rubble-related appeals. These hotspots should not be interpreted as a complete or independently verified map of structural damage
Manual validation showed that human reviewers confirmed usable location data in 589 of 673 reviewed posts, while 84 posts had non-verifiable location data. The algorithm missed 74 of the 589 usable location posts and falsely detected location data in 31 of the 84 non-verifiable posts. The location extraction achieved a precision of 94.3%, recall of 87.4%, F1-score of 90.7%, overall accuracy of 84.4%, and specificity of 63.1% (Table 5).
Table 5.
Manual validation of usable location detection
| Measure | Result |
|---|---|
| Manually reviewed posts | 673 |
| Human confirmed usable location posts | 589 |
| Posts with no verifiable location data | 84 |
| Missed usable location posts | 74 |
| False positive location detections | 31 |
| Precision | 94.3% |
| Recall / Sensitivity | 87.4% |
| F1 score | 90.7% |
| Overall accuracy | 84.4% |
| Specificity | 63.1% |
Discussion
Our analysis indicates that social media engagement after the earthquake progressed through certain phases, reflecting the evolving needs of society.
Initial rescue-appeal period (first 12 hours)
The posts sent in the first hours of the earthquake (starting at 04:17 TRT) were characterized by shock and urgent survival appeals. These messages included statements such as “earthquake occurred” and expressions of shock such as “Oh my God”, “get well soon”, which indicate an immediate cognitive response to the event. This content provides insight into the mental health of the affected population [9]. Between 8 and 9 a.m., “I am under rubble”, “still under rubble” (n = 229.1 thousand) emerged as the highest-ranked expressions, excluding the general “earthquake” hashtag. This indicated a shift to an intense period of crowdsourced rescue appeals. Highly specific location data including building names (e.g., Arzu apartment) and district names (e.g., Odabaşı) were noticed in the following hours. Frequent use of regional district names during early critical hours may indicate decentralized and highly localized information bottlenecks that formal reporting systems can ignore. This level of detail may be useful for search-and-rescue coordination if the information can be quickly verified, geocoded, and integrated with official response systems. The increase in requests for external aid after 8:00 a.m. demonstrated an early public awareness of the destruction and lack of local capacity.
Logistics and shelter period (days 2–4)
Following the first 24 hours, the attention shifted to critical humanitarian logistics (e.g., take water to region, need for shrouds). Despite the ongoing requests for rescue during the second and third days, new hashtags appeared. This represents emerging priorities such as shelter and housing, drinkable water, medications, cranes, and mobile toilets.
A key finding within the “Requests and Needs” category was the difference between volume and engagement. Requests for blankets had the highest volume (n = 43,582). However, despite its lower volume (n = 2,257), tent-related posts received more engagement (180.01 retweets per post). The high level of engagement with tent-related posts suggests that shelter was highly salient to users and was perceived as an urgent concern within the online discourse [28]. It also indicates that tents were viewed as both harder to access and more urgently needed. These findings are consistent with the post-disaster assessment of 2023 Türkiye-Syria earthquakes, which found providing shelter more challenging than delivering other aid [29]. Similarly, the media’s depiction of public outrage concerning the insufficient number of tents and difficulties in their distribution, further corroborates this issue. High retweet rates for cranes and tents suggest that posts about specific, actionable resources received greater public amplification than more general humanitarian messages.
Socio-economic and recovery period (first week)
In the following days, X functioned as a platform for citizen engagement, policy advocacy, and socio-economic recovery. The topics expanded from immediate survival to encompass demands for the cancellation of stock market transactions, as well as solutions for distance education, and temporary accommodation. This reflects how disasters trigger the restructuring of public priorities such as economic continuity, social cohesion, and stability [30]. The delayed appearance of specific public health requests, such as for “sanitary pads” and “baby diapers/rash cream”, indicates a gap in addressing certain specialized needs. This pattern suggests that X may provide supplementary information about emerging public concerns that could inform policy discussions when interpreted alongside verified field assessments. These findings are relevant both to emergency response and to understanding the longer-term socioeconomic impacts of the disaster [7].
Digital signals of public health needs during disaster response
The analysis of early content highlights four prominent types of disaster-related health priorities visible in the dataset. The first priority was search and rescue, reflected in location-specific posts such as “I’m under rubble” and “voice from the rubble”, as well as messages including addresses and street numbers. Several studies have indicated that X can function as an improvised tool for crowd-sourced mapping of incident sites [10]. In the dataset, phrases like “I’m under rubble,” “help,” or specific building addresses contain direct, real-time rescue related information.
Following the earthquake, the “deprem.io” platform was launched as a community-driven project to assess rubble location data acquired from X. However, the analysis of non-geotagged posts uncovered a significant limitation: spatial information supplied by victims on X cannot be automatically retrieved and converted into accurate and machine-readable structures. Enhancing the clarity and structure of such information could improve the usability of crowdsourced rescue-related information with current damage assessment technologies [31].
A second priority emerged for emotional support. Tweets expressing shock such as “Oh my God”, and prayer messages “get well soon” indicated an urgent need for psychological aid. Emotional intensity conveyed through this content reflects a collective anxiety following the disaster. This demonstrates the necessity of deploying pre-established mental health communication strategies not only to provide assistance but also to mitigate the spread of distress and confusion on digital platforms [32].
The third priority involved the requests for essential resources. Requests for items like “blankets”, “take water to region”, and “pharmacies” highlighted acute logistic challenges, particularly the importance of pre-positioning high-demand non-food items near high-risk areas. Our results align with existing research on humanitarian logistics, which shows the value and efficiency of forward-locating resources during peak demand [33]. In this context, real-time social media signals may provide supplementary signals about localized needs when combined with verified field assessments. Finally, a fourth and later emerging priority focused on hygiene and sanitation needs. Posts calling for “mobile toilets”, “sanitary pads”, and “masks” revealed a gap in how specific health and hygiene needs were addressed during early response. These signals reflect more than logistical delays; they indicate structural oversights in preparedness planning [31, 33]. These findings extend previous crisis informatics research by showing how specific public health issues, especially hygiene, sanitation, and dignity-related requirements, were increasingly prominent over time in this earthquake dataset. Because the present analysis is descriptive, the observed patterns should be interpreted as indicators of online discourse and public amplification, rather than as validated measures of field-level need, response effectiveness, or resource scarcity.
Engagement metrics and public amplification of disaster needs
Virality, measured by average retweets and likes, was used as an indicator of public amplification and engagement during crises [4]. Under high stress and urgency, users tend to share messages that offer clear, practical value [21]. However, these indicators are prone to artificial inflation by automated bot networks and coordinated amplification efforts, and should not be treated as a direct measure of need severity, logistical scarcity, or operational priority without external validation. According to the results, tent aid requests were retweeted nearly 15 times more often (average retweets of 180.01 per tweet) than the high-volume requests for blankets (average retweets of 12.10 per tweet). These engagement patterns may help generate hypotheses about which needs were most visible or urgent within the online public sphere. Tents represent complex, scalable shelter solutions that are critically scarce, while blankets are easily sourced items. The high amplification of tent requests suggests that shelter-related needs were especially visible and widely circulated among X users, although this visibility cannot be assumed to reflect purely organic prioritization by users. High virality may reflect moments when users collectively amplified specific needs [3]. These findings suggest that preparedness policies could consider monitoring engagement metrics as one supplementary input, provided that such signals are verified and combined with official needs assessments [5].
Public trust appears to be an important determinant of message spread [6]. Hashtags related to trusted organizations and figures (e.g., Turkish Red Crescent, Ahbap Association, Haluk Levent) received significant engagement (89.62 average retweets for the category), contrasting with low engagement for symbolic, propaganda-oriented hashtags such as “85 million one heart” and “state and nation together” with average retweets of only 8.79 and 22.51, respectively. This underscores that during a crisis, the public pays greater attention to credible information from the actors over symbolic campaigns [11]. For government officials, these findings provide feedback: messages should focus on verified aid delivery and transparent logistics, not symbolic campaigns. Table 6 outlines how content type, trust, and audience engagement are linked.
Table 6.
Social media engagement metrics and disaster resilience strategy
| Content Category | Observed Engagement (Retweet Avg.) | Resilience/Preparedness Strategy | Crisis Communication Implication |
|---|---|---|---|
| Humanitarian Support, Services | High average retweets (89.62) | Recognizing that posts associated with humanitarian actors may receive high public engagement and could be included in communication monitoring frameworks. | Trust is built on action; institutions associated with direct aid receive immediate public amplification. |
| Requests and Needs (Tent Aid) | Extremely High Retweets (180.01) | Considering engagement metrics as supplementary signals for identifying highly visible needs, subject to verification with field data (highlighting highly visible resource related concerns). | Posts describing specific needs may receive stronger public amplification than general expressions of support. |
| Propaganda & Campaign | Low average retweets (8.79) | Developing authenticity in government messaging; shifting focus from symbolic unity to verifiable, practical assistance delivery. | Symbolic or general campaign messages received lower engagement than some need-specific and humanitarian-service messages in this dataset. |
| Policy & Long-Term Needs | Qualitatively observed via trending topics (Table 3) | Utilizing social media intelligence for long-term recovery planning and assessing socio-economic pressure points. | Indicates that X discourse in the dataset expanded from immediate rescue concerns to recovery, education, economic, and governance-related topics. |
Managing misinformation and digital resilience
While virality facilitates the rapid spread of urgent help requests, it also amplifies the risk of spreading misinformation. Evidence suggests that misinformation spreads significantly faster than true news, often via human actors rather than bots [34]. Policies on managing information must therefore address the challenge of optimizing the dissemination of requests while implementing enhanced verification for highly shared content. A practical starting point would involve training citizens and first responders to provide verifiable, machine-readable data from the outset, allowing for swifter authentication of high-priority requests.
This highlights a critical limitation: the inability to automatically extract reliable and machine-readable spatial data from tweets, which necessitated a reliance on manual validation. Because the success of rescue operations depends on the speed of location data [29, 35], there is a clear need for increased investment in digital infrastructure and automated tools [12]. Technology that can immediately translate a non-geotagged report into a GIS coordinate would transform social media from a general “signal” into a precise operational tool. Moreover, the real-time monitoring capability of X, shown by the correlation between seismic events and posting frequency, could be explored as a supplementary communication input within broader early warning and response systems [36]. The visibility of trusted scientific figures in trending topics also shows a public demand for technical information. Policies should establish direct communication channels for these experts to utilize their credibility in crisis management.
Beyond the immediate response, social media intelligence remains a vital tool during the recovery phase for tracking how public priorities change over time. In the later stages of the disaster, public discourse shifted noticeably toward issues of governance and financial stability. This result is in line with research reflecting public anxiety regarding economic losses and long-term uncertainty [37]. By monitoring policy-related hashtags, authorities can gain a more responsive understanding of public sentiment as it unfolds [38]. Consequently, preparedness plans should include economic intervention strategies designed to address financial panic and support social stability throughout the extended recovery period [39].
Limitations
Several limitations should be considered when interpreting the findings of the study. First, the technical constraints of the X API v2 search endpoint and keyword-based retrieval strategy set restrictions on data collection. The API retrieves only posts that were publicly available at the exact moment of data extraction, as it functions as a temporal cross-sectional sample. Any tweets that were deleted by users or removed by moderators prior to extraction are systematically absent. Furthermore, reliance on predefined Boolean queries means the dataset may exclude relevant discourse that utilized alternative phrasing, typographical errors, or implicit references. In addition, the category-level engagement analysis was limited to selected high-frequency hashtags and phrases. To ensure categorical clarity and prevent double-counting, we excluded posts that either lacked specific keywords or contained overlapping hashtags. While this approach likely underrepresents less frequent, mixed-topic, or more complex posts, it maintains the integrity of the individual categories. Consequently, the temporal peaks and engagement metrics observed in the study should be interpreted as patterns within the analyzed dataset rather than exact estimates of total X activity.
Second, by focusing specifically on original posts, our methodology excludes the broader spectrum of engagement involving NGOs and government organizations on X. This gap is further widened by the relative inactivity of these organizations on X, which likely skewed the representation of their institutional efforts within the final dataset. Beyond these organizational limits, the reliance on social media data introduces inherent demographic and geographic imbalances. Because individuals in rural regions or those with lower digital literacy are often marginalized in digital datasets, the findings primarily reflect the perspectives of users with consistent internet access. Consequently, while the data captures a specific segment of the discourse, it may overlook some of the most severely affected and hard-to-reach communities.
Third, individual posts were not cross-referenced with official records regarding rescue, damage, or logistics. While engagement metrics like retweets and likes provide a clear picture of what the public amplified, they do not necessarily correlate with the actual severity of a crisis or the effectiveness of a response. This distinction is important because engagement counts on X can be skewed by automated bots or coordinated campaigns rather than reflecting organic public concern. Consequently, the metrics in this study should be viewed as measures of platform visibility rather than absolute indicators of public priority. Similarly, the timing of these posts shows how needs were expressed online, but this does not imply that social media data can reliably forecast future resource needs. To build on this work, future research should integrate social media analysis with bot detection tools and official field reports to better assess data accuracy and operational value.
Finally, the heatmaps are subject to spatial uncertainty. The precision of these visualizations depends on a combination of factors, ranging from the reliability of user-provided details to the technical performance of our extraction and geocoding workflows. For example, incomplete or inaccurate data can compromise spatial accuracy, while the deduplication of reports might inadvertently mask the true density of distinct collapse sites located in close proximity. Although manual validation showed satisfactory performance performance for usable location detection, the validation sample was not stratified by province, date, or time period. Moreover, the study did not independently verify the field accuracy of every geocoded coordinate. These heatmaps should be viewed as approximate representations of digital rescue requests rather than definitive records of structural damage or operational field activity. Future research should also investigate how cultural, regional, and technological factors shape digital engagement during disasters.
Conclusion
This research contributes to established crisis-informatics literature by providing a detailed analysis of Turkish-language disaster communication during the 2023 earthquakes. A key finding is the documentation of how public priorities evolved on social media; while early discourse was dominated by urgent rescue requests, the focus steadily shifted toward longer-term recovery needs such as shelter, hygiene, and sanitation. Findings confirm that X was actively used in both disseminating urgent calls for assistance and in the monitoring of public health and logistics issues.
The analysis supported three conclusions. First, X served as a decentralized channel through which users shared urgent, location-specific rescue appeals during the initial hours of the disaster. Second, the findings show that tweet volume and engagement metrics captured different dimensions of online attention. While high-volume terms reflected widely expressed needs, highly retweeted posts indicated which needs received stronger public amplification. These metrics should be interpreted as descriptive indicators of online salience rather than validated measures of logistical bottlenecks. This was evidenced by the significantly higher engagement for tents compared to more general requests for blankets. Third, public trust is paramount in crisis communication. The data consistently showed higher engagement with messages from trusted civil organizations and those offering direct aid. In contrast, symbolic or non-actionable campaigns received significantly less attention.
These findings suggest several considerations for disaster communication and public health preparedness. Recommendations include investing in digital infrastructure, especially automated geoparsing tools to generate usable GIS coordinates. Humanitarian and government organizations could explore the use of engagement metrics as supplementary monitoring signals, provided that they are combined with verification procedures, official field reports, and safeguards against misinformation. To address these challenges effectively, official communication must prioritize both clarity and practical utility. Beyond the messaging itself, public health teams should incorporate sentiment monitoring into their workflows to identify and mitigate the rapid escalation of public distress. Strengthening disaster preparedness further requires the strategic integration of scientific experts and civil society organizations into official networks, thereby fostering a more transparent and reciprocal dialogue between institutions and the public.
Acknowledgements
The authors would like to thank Prof. Figen Ünal Çolak for her valuable insights and discussions on the topic of this research.
Authors' contributions
M.F.E.: Conceptualization (lead); methodology (lead); data curation (lead); writing - original draft (equal); formal analysis (lead); software (lead), visualization (lead), writing - review and editing (equal); final approval of the version to be submitted (equal). F.K.Y: Conceptualization (supporting); writing - original draft (equal); writing - review and editing (equal); final approval of the version to be submitted (equal). S.K.: Writing - original draft (supporting); final approval of the version to be submitted (equal).
Funding
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Data availability
The data that support the findings of this study are available from the corresponding author upon request.
Declarations
Ethics approval and consent to participate
The authors confirm that ethical approval was not required for this study as it does not involve human participants or animal subjects.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.GFDRR - Global Facility for Disaster Reduction and Recovery. February 2023 Kahramanmaraş, Türkiye, Earthquakes. Available from: https://www.gfdrr.org/en/publication/grade-report-february-2023-kahramanmaras-turkiye-earthquakes.
- 2.TÜRKONFED. Afet durumu: İki büyük deprem. 2023. Available from: https://turkonfed.org/Files/ContentFile/turkonfed2023kahramanmarasdepremiafetdurumraporu021023-4718.pdf.
- 3.Figueiredo F, Almeida JM, Gonçalves MA, Benevenuto F. On the dynamics of social media popularity. ACM Trans Internet Technol. 2014;14(4):1–23. 10.1145/2665065. [Google Scholar]
- 4.Noor N, Okhai R, Jamal TB, Kapucu N, Ge YG, Hasan S. Social-media-based crisis communication: assessing the engagement of local agencies in Twitter during Hurricane Irma. Int J Inform Manage Data Insights. 2024;4(2):100236. 10.1016/j.jjimei.2024.100236. [Google Scholar]
- 5.Karimiziarani M, Jafarzadegan K, Abbaszadeh P, Shao W, Moradkhani H. Hazard risk awareness and disaster management: extracting the information content of twitter data. Sustainable Cities Soc. 2021;77:103577. . 10.1016/j.scs.2021.103577 [Google Scholar]
- 6.DiCarlo MF, Berglund EZ. Connected communities improve hazard response: an agent-based model of social media behaviors during hurricanes. Sustainable Cities Soc. 2021;69:102836. 10.1016/j.scs.2021.102836. [Google Scholar]
- 7.Ogie R, Moore A, Wickramasuriya R, et al. Twitter data from the 2019–20 Australian bushfires reveals participatory and temporal variations in social media use for disaster recovery. Sci Rep. 2022;12:16914. 10.1038/s41598-022-21265-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Statista. Number of social network users worldwide 2017–2030. 2025 Nov 10. Available from: https://www.statista.com/forecasts/278414/number-of-worldwide-social-network-users.
- 9.Teague S, Shatte A, Weller E, Fuller-Tyszkiewicz M, Hutchinson D. Methods and applications of social media monitoring of mental health during disasters: scoping review. JMIR Ment Health. 2022;9(2):e33058. 10.2196/33058. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Imran M, Castillo C, Diaz F, Vieweg S. Processing social media messages in mass emergency: a survey. ACM-CSUR. 2015;47(4):1–38. 10.1145/2771588. [Google Scholar]
- 11.Eriksson M. Lessons for crisis communication on social media: a systematic review of what research tells the practice. Int J Strategic Communication. 2018;12(5):526–51. 10.1080/1553118X.2018.1510405. [Google Scholar]
- 12.Rachunok B, Fan C, Lee R, Nateghi R, Mostafavi A. Is the data suitable? The comparison of keyword versus location filters in crisis informatics using Twitter data. Int J Inform Manage Data Insights. 2022;2(1):100063. 10.1016/j.jjimei.2022.100063. [Google Scholar]
- 13.Kinsky ES, Chen L, Drumheller K. Crisis and emergency risk communication: FEMA’s Twitter use during the 2017 hurricane season. Public Relations Rev. 2021;47(4):102094. 10.1016/j.pubrev.2021.102094. [Google Scholar]
- 14.Kankanamge N, Yigitcanlar T, Goonetilleke A, Kamruzzaman M. Determining disaster severity through social media analysis: testing the methodology with South East Queensland Flood tweets. Int J Disaster Risk Reduct. 2019;42:101360. . 10.1016/j.ijdrr.2019.101360 [Google Scholar]
- 15.Martínez-Rojas M, Del Pardo-Ferreira C, Rubio-Romero M. Twitter as a tool for the management and analysis of emergency situations: a systematic literature review. Int J Inf Manag. 2018;43:196–208. 10.1016/j.ijinfomgt.2018.07.008. [Google Scholar]
- 16.Aladwani AM. Facilitators, characteristics, and impacts of Twitter use: theoretical analysis and empirical illustration. Int J Inf Manag. 2014;35(1):15–25. . 10.1016/j.ijinfomgt.2014.09.003 [Google Scholar]
- 17.Kitazawa K, Hale SA. Social media and early warning systems for natural disasters: a case study of Typhoon Etau in Japan. Int J Disaster Risk Reduct. 2020;52:101926. . 10.1016/j.ijdrr.2020.101926 [Google Scholar]
- 18.Landwehr PM, Wei W, Kowalchuck M, Carley KM. Using tweets to support disaster planning, warning and response. Saf Sci. 2016;90:33–47. 10.1016/j.ssci.2016.04.012. [Google Scholar]
- 19.Vicente P. Sampling Twitter users for social science research: evidence from a systematic review of the literature. Qual Quant. 2023;57:5449–89. 10.1007/s11135-023-01615-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Murthy BP, LeBlanc TT, Vagi SJ, Avchen RN, Going Viral. The 3 Rs of Social Media Messaging during Public Health Emergencies. Health Secur. 2021;19(1):75–81. 10.1089/hs.2020.0157. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Pourebrahim N, Sultana S, Edwards J, Gochanour A, Mohanty S. Understanding communication dynamics on Twitter during natural disasters: a case study of Hurricane Sandy. Int J Disaster Risk Reduct. 2019;37:101176. 10.1016/j.ijdrr.2019.101176. [Google Scholar]
- 22.Hilberts S, Govers M, Petelos E, Evers S. The impact of misinformation on social media in the context of natural disasters: narrative review. JMIR Infodemiology. 2025;5:e70413. 10.2196/70413. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Omar S, Van Belle JP. Disaster misinformation management: strategies for mitigating the effects of fake news on emergency response. In: Rocha Á, Ferrás C, Hochstetter Diez J, Diéguez Rebolledo M, editors. Information Technology and Systems. ICITS 2024. Lecture Notes in Networks and Systems, vol 932. Cham: Springer; 2024. Available from: 10.1007/978-3-031-54235-0_29.
- 24.Ginzarly M, Teller J, Dujardin S. Social media use in disaster response: empowering community resilience. Int J Digit Earth. 2025;18(1). 10.1080/17538947.2025.2521791.
- 25.Chari F, Ngcamu BS. Social media and inclusive humanitarian intervention: the case of Cyclone Idai in Chimanimani district, Zimbabwe. J Humanitarian Logistics Supply Chain Manage. 2024;15(2):107–22. 10.1108/jhlscm-05-2023-0047. [Google Scholar]
- 26.Paulus D, Fathi R, Fiedrich F, et al. On the interplay of data and cognitive bias in crisis information management. Inf Syst Front. 2024;26:391–415. 10.1007/s10796-022-10241-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Lai CH, Tang T. Information exposure and information overload as antecedents of crisis communicative responses and coping: a cross-country comparison. Am Behav Sci. 2025;69(12):1439–61. 10.1177/00027642221132805. [Google Scholar]
- 28.Spearing LA, Stephens KK, Faust KM. Shelter shopping: where the built environment and social systems meet. Int J Disaster Risk Reduct. 2021;102161. 10.1016/j.ijdrr.2021.102161.
- 29.Sarı B. Disaster management following the great Kahramanmaraş earthquakes in 2023, Türkiye. Nat Hazards Earth Syst Sci. 2025;25(6):2031–43. 10.5194/nhess-25-2031-2025. [Google Scholar]
- 30.He C, Hu D. Social media analytics for disaster response: classification and geospatial visualization framework. Appl Sci. 2025;15(8):4330. 10.3390/app15084330. [Google Scholar]
- 31.Yuan F, Liu R. Feasibility study of using crowdsourcing to identify critical affected areas for rapid damage assessment: Hurricane Matthew case study. Int J Disaster Risk Reduct. 2018;28:758–67. 10.1016/j.ijdrr.2018.02.003. [Google Scholar]
- 32.García YE, Villa-Pérez ME, Li K, Tai XH, Trejo LA, Daza-Torres ML, et al. Wildfires and social media discourse: exploring mental health and emotional wellbeing through Twitter. Front Public Health. 2024;12:1349609. 10.3389/fpubh.2024.1349609. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Shariati A, L’Hermitte C, Trent NM. Decision-making insights on the prepositioning of relief items: a systematic review. J Humanitarian Logistics Supply Chain Manage. 2026;16(1):70–88. 10.1108/JHLSCM-07-2024-0091. [Google Scholar]
- 34.Vosoughi S, Roy D, Aral S. The spread of true and false news online. Science. 2018;359(6380):1146–51. 10.1126/science.aap9559. [DOI] [PubMed] [Google Scholar]
- 35.Koshy R, Elango S. Applying social media in emergency response: an attention-based bidirectional deep learning system for location reference recognition in disaster tweets. Appl Intell. 2024;54:5768–93. 10.1007/s10489-024-05462-6. [Google Scholar]
- 36.Contreras D, Wilkinson S, Alterman E, et al. Accuracy of a pre-trained sentiment analysis (SA) classification model on tweets related to emergency response and early recovery assessment: the case of 2019 Albanian earthquake. Nat Hazards. 2022;113:403–21. 10.1007/s11069-022-05307-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Ma X, Xia S, Ding L. Building community resilience and calling for collective actions: how corporations and publics communicate disaster aid on social media. Humanit Soc Sci Commun. 2025;12:726. 10.1057/s41599-025-05049-7. [Google Scholar]
- 38.Chandrasekaran R, Mehta V, Valkunde T, Moustakas E. Topics, trends, and sentiments of tweets about the COVID-19 pandemic: temporal infoveillance study. J Med Internet Res. 2020;22(10):e22624. 10.2196/22624. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Aksoy CG, Chupilkin M, Koczan Z, Plekhanov A. Unearthing the impact of earthquakes: a review of economic and social consequences. J Policy Anal Manag. 2025;44:1450–71. 10.1002/pam.22642. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon request.