When the July 2021 floods struck western Germany, Instagram filled with images of submerged streets, rescue boats and volunteers shoveling mud. A new study has now read roughly 2,700 of those posts with the help of generative artificial intelligence, and it reports that the visual story people told shifted over three years from destruction toward solidarity, recovery and remembrance, while the emotional tone moved from despair and anxiety toward steadier expressions of hope.
The research, published in Computational Urban Science, focused on the extreme flooding that hit western Germany in July 2021, an event that affected more than 107,000 people. The researcher collected Instagram posts tagged with three widely used hashtags, #germanyflood, #germanyfloods and #hochwasser2021, between July 2021 and July 2024. After removing duplicates from about 5,000 scraped photos, the final dataset contained 2,690 images, each stored with metadata including the image URL, user name, date and hashtags.
The method combined generative AI with qualitative and quantitative analysis. Images were interpreted by ChatGPT-4 under human supervision through a pre-defined prompt that asked for descriptions of the setting, visible objects, flood-related damage, human presence and actions. The model processed images in small batches, because testing showed that uploading more than 10 images at a time caused errors, likely due to input limits. The prompt first asked for a general description of every visible element, including any text shown in signs or overlays, and then for two binary judgments: whether the image was related to flooding and whether people appeared in it.
Validation was a central part of the design. The researcher manually coded a random 30 percent sample of the photos, and six main themes emerged from that inductive coding. ChatGPT-4 was then instructed to apply the same coding logic to the remaining 70 percent. Early results were uneven: binary coding of people and damage reached 90 percent accuracy, but the AI initially placed nearly 70 percent of images under flood damage, letting the pervasive presence of destruction overshadow less salient themes such as coordination, solidarity or recovery. One image described as showing residential buildings partially submerged with rescue workers in a boat was first labeled flood damage even though rescue activity was visible. Additional decision rules, including an instruction to choose the label that best represents the main focus of an image, corrected this imbalance.
The refined AI-assisted coding was checked against human annotations and showed substantial agreement, with an overall percentage agreement of 80.73 percent and a Cohen’s kappa of 0.74. Category-level performance was strong for several themes, with F1 scores of 0.88 for solidarity and resilience, 0.84 for rescue operations and relief efforts, and 0.83 for flood damage. The six themes that structured the analysis were flood damage, rescue operations and relief efforts, solidarity and resilience, media coverage and public awareness, post-disaster recovery, and government and official involvement.
The most prominent theme across the corpus was solidarity and resilience, with 968 occurrences, or 36 percent of posts, followed by flood damage with 633 occurrences, 23.5 percent, and rescue operations and relief efforts with 496 occurrences, 18.4 percent. Post-disaster recovery accounted for 335 occurrences, media coverage and public awareness for 223, and government and official involvement for 35. The author notes that this distribution differs from existing automated classification studies, which tend to emphasize physical damage, and that solidarity and resilience emerged as the dominant visual narrative, highlighting collective action and volunteerism often overlooked in conventional disaster image classification frameworks.
The temporal distribution of themes traced a movement from immediate impacts toward reconstruction and remembrance. Posting peaked in the weeks between July 12 and July 25, 2021, with 898 unique posts, and the year 2021 produced 1,869 unique posts, followed by 487 in 2022, 229 in 2023 and 105 in 2024. In July 2021 the dominant themes were flood damage, solidarity and resilience, and rescue operations. By August, representations of solidarity and relief operations became more prominent than flood damage, suggesting an early turn from documenting destruction toward mutual aid and volunteer mobilization. In 2022, solidarity and resilience remained relatively stable with a pronounced peak in July, coinciding with the anniversary of the flood, while post-disaster recovery grew gradually in the second half of the year, with peaks in July, August and September.
The emotional tone of the images was inferred by ChatGPT from facial expressions and the textual descriptions produced during visual analysis, and the pilot phase showed why refinement was needed: the first 30 posts yielded 27 distinct emotional labels, so iterative prompting grouped closely related expressions into broader categories until a final taxonomy of seven broad emotional categories was reached, with one dominant emotion assigned per post. Overall, 59 percent of the images featured people. From July to November 2021, hope was the most prevalent tone with 431 occurrences, followed by neutral with 317, despair with 277, sadness with 231 and anxiety with 214. Despair peaked in July 2021 and gradually declined in the following weeks, while hope remained consistently present and the tone became more neutral as the event progressed. A chi-square test found a statistically significant association between visual themes and emotional tones, with chi-square of 460.82 on 70 degrees of freedom and p below 0.001. Flood damage was primarily associated with despair, rescue operations with anxiety, solidarity and resilience combined hope, neutrality and joy, post-disaster recovery was dominated by hope and joy, and media coverage was associated with sadness and seriousness.
A complementary thematic analysis of hashtags showed how public discourse around the disaster was framed. After hashtags directly tied to the event, the most frequent were those associated with community resilience and solidarity, such as #strongertogether, #wekeepgoing and #weAHRfamily, followed by location-specific tags like #ahrtal, then relief-related tags, climate-related tags and media coverage tags. Because a single post can carry hashtags from multiple themes, hashtag frequency does not directly reflect theme frequency within posts: 603 posts included solidarity and resilience hashtags and 980 fell under rescue and relief efforts, with an overlap of 582 posts, or 21.6 percent, showing both themes at once. Climate-related hashtags appeared in only 197 posts, and #climatechange appeared 140 times, about 5.2 percent of posts, a relatively low share that contrasts with Twitter studies of the same floods in which climate change discussions ranked among the most prevalent topics, a difference the study attributes to Instagram being primarily photo-oriented while Twitter favors brief, hashtag-driven messages.
The study also documents who posted. A large portion of the content came from unique users, with 958 individual users posting once, 162 posting twice and 61 posting three times, indicating a highly distributed pattern of engagement, while a smaller group of 70 accounts posted between four and ten times. One nonprofit flood relief organization, DZN Hilfsorganisation, accounted for 341 posts, followed by the regional news outlet SWR Aktuell with 140. The author cautions that the corpus is bounded by the selected hashtags and by the platform’s API functionality, algorithmic filtering and datafication, so the observed decline in tagged posts reflects activity associated with the event-related hashtags used for retrieval rather than a direct measure of overall community engagement, and a growing proportion of relevant content may remain uncaptured as users shift to hyperlocal tags or private channels. Within those limits, the analysis demonstrates how AI-assisted visual interpretation, combined with human supervision, can make large-scale analysis of social media image corpora feasible and generate insights relevant to disaster response and recovery strategies, from real-time information during the response phase to evidence for policymakers on rehabilitation and restoration efforts during recovery.
Subject of Research: AI-assisted multimodal analysis of Instagram posts documenting the July 2021 western Germany floods across response and recovery phases
Article Title: Visual narratives of disaster on Instagram: a multimodal AI-assisted analysis
Article References: Ginzarly, M. (2026). Visual narratives of disaster on Instagram: a multimodal AI-assisted analysis. Computational Urban Science, 6(1), Article 63. https://doi.org/10.1007/s43762-026-00300-6
Image Credits: AI Generated
DOI: 10.1007/s43762-026-00300-6
Keywords: Instagram, floods, Germany 2021, generative AI, ChatGPT-4, disaster response, disaster recovery, social media, visual analysis, emotional tone, solidarity, hashtags
News Source: Violet Maxwell. (October 28, 2026). AI reads thousands of Instagram flood photos and finds hope after despair. Scienmag.
