Scientists Mine Instagram Photos to Map a Treeline Species Across 2,000 Kilometers of the Andes

Every autumn, hikers in Patagonia photograph the same spectacle: entire mountainsides of lenga beech trees glowing orange and red beneath snow-capped peaks. Those vacation snapshots, mostly taken by people with no interest in science, have now become an unlikely scientific resource. In a study published in Frontiers of Biogeography, researchers describe a novel ground truthing approach that harvests geotagged Instagram posts to validate large-scale maps of Nothofagus pumilio, the dominant treeline species of the southern Andes. By systematically searching hashtags and location tags, the team compiled 1,238 occurrence points spanning roughly 2,000 kilometers of latitude, from 35°S to the southern tip of Tierra del Fuego, and used them to check a satellite-derived distribution map of the species.

The motivation behind the study is a familiar problem in biodiversity research. Species distribution models are only as good as the occurrence data feeding them, and field sampling in remote high-elevation terrain is expensive, slow, and often physically dangerous. As a shortcut, ecologists frequently download records from open databases such as the Global Biodiversity Information Facility, or GBIF, which aggregates observations from museums, herbaria, and citizen science platforms. Yet such data are riddled with hidden biases: observers cluster near cities, roads, and tourist attractions, while wilderness areas and rugged slopes remain blank on the map. For the treelines of the southern Andes, the problem is particularly acute. Compared with the well-studied mountains of the Northern Hemisphere, Patagonian treelines have seen mostly small, local studies, and large-scale modeling work has been hampered by the simple fact that treeline sites are extremely hard to reach.

The researchers’ solution exploits an unusual combination of Instagram’s features. Unlike text-dominated platforms, Instagram hosts only photos and videos, and it has roughly two billion users worldwide, many of them visiting exactly the national parks where lenga beech grows. Users typically attach geographic tags, location descriptions, and hashtags to their posts, which allows each image to be traced to a place on the map. The team began in 2021 by searching the hashtags #nothofaguspumilio and #lenga, then systematically searched for named locations across the species’ range to fill gaps, aiming for a homogeneous spread of points. Each candidate post was screened against a strict catalogue of criteria before it could contribute data.

Those criteria were demanding. The tree had to be clearly identifiable, either through morphological features such as its elliptical, notched leaves, or through the species’ most reliable signature: the vivid autumn colouring that distinguishes the deciduous lenga from the evergreen Nothofagus species sharing its range. The location had to be traceable not just from a caption but from landscape features visible in both the photograph and a satellite image, such as glaciers, distinctive peaks, lakes, roads, or coastlines. Images centered on people, montages, emoji-laden posts, and heavily filtered or colour-distorted pictures were excluded. Where a post qualified, the researchers placed a point at the photo location, and often additional points where the autumn-coloured treeline was visible in the background of landscape shots. Points were assigned to a one-kilometer grid, a common target resolution for ecological models. In total, 297 suitable posts published between 2017 and 2022 yielded the 1,238 ground truthing points.

The resulting dataset was then benchmarked against GBIF. After filtering the database for coordinate quality, just 558 usable records of N. pumilio remained, most contributed through the citizen science platform iNaturalist, spanning sampling years from 1981 to 2024. The comparison was revealing. Nearly half of the GBIF points, 252 in total, were concentrated in a handful of tourist and urban hotspots: 159 points around Ushuaia, 53 in Torres del Paine National Park, 24 around Bariloche, and 16 at a single parking area with a viewing platform for the Perito Moreno Glacier. A quantitative bias analysis using the R package sampbias confirmed that cities exerted a significantly stronger pull on the GBIF dataset than on the Instagram-derived points, which were also more evenly split between the northern and southern halves of the study area. The Instagram dataset was not bias-free, showing its own clustering around lakes in famous parks, but the structured search strategy clearly reduced the skew toward human-accessible sites. On average, Instagram points sat at 1,049 meters elevation, reflecting the deliberate preference for posts showing the treeline itself, whereas the filtered GBIF records averaged about 560 meters, mostly from the species’ more accessible southern range.

With ground truth in hand, the team turned to the satellite side of the project. They downloaded Sentinel-2 level 2A imagery at 20-meter resolution for summer and autumn seasons between 2019 and 2022, masked non-vegetated pixels using the sensor’s Scene Classification Layer, and ran a supervised classification separating lenga beech, evergreen vegetation, and low vegetation. The species’ ecology worked in the researchers’ favor: lenga forms near-pure stands at the treeline, and its autumn senescence makes it spectrally and visually distinct. The best-performing classifiers, including a Random Forest approach, achieved overall accuracies of 0.93 for the summer scenes and 0.97 for the autumn scenes, with matching Kappa values of 0.89 and 0.96. Elevation thresholds drawn from the literature were applied in the north of the range to strip out deciduous lowland species misclassified as lenga. The final map captured the species from 35°10’S to 55°59’S, including the characteristic upper treeline and, in the arid east, a second lower, xeric treeline toward the forest-steppe ecotone.

Validation delivered a striking asymmetry. Of the 1,238 Instagram ground truthing points, 1,142, or 92.25 percent, fell within areas the classification had mapped as lenga beech. The 96 disagreements clustered in areas plagued by mountain shadows and data gaps caused by persistent cloud cover at the southern tip of the continent, where entire Sentinel-2 scenes were unavailable or unusable during the sensing period. By contrast, only 157 of the 558 GBIF points, about 28 percent, aligned with the mapped distribution. The authors caution that many of these mismatches lie just outside the classified area and likely reflect coordinate uncertainty, records taken on trails beside the forest rather than within it, or isolated ornamental and urban trees that a landscape-scale classifier cannot capture. Still, the gap between 92 percent and 28 percent underlines how much precision is lost when records carry coordinate errors of up to a kilometer, a substantial deviation in complex mountain terrain.

The method comes with honest limitations. The manual search of posts and the hand placement of points on maps is labor-intensive and slow, and unlike Flickr-based studies that have used automated queries through application programming interfaces, the Instagram API is difficult to access, making manual screening unavoidable for now. Publication dates on Instagram also do not guarantee when a photograph was actually taken, a problem for any analysis of temporal change, and historical observations simply do not exist on the platform. There are ecological ambiguities too: in parts of the range, the closely related deciduous N. antarctica, which can hybridize with lenga, proved indistinguishable in the satellite classification. The researchers also engage directly with the ethics of repurposing social media content, noting that content creators are rarely aware of such reuse. To address this, the team created a public Instagram account presenting the study and its results, liked every image they used, asked users for permission where posts appeared in publications, and used Instagram’s chat function to exchange information with photographers about locations.

The broader promise is considerable. The authors argue that the Instagram ground truthing approach could be transferred to other visually distinctive species, provided that landscape elements appear in posts, which favors landscape photography over close-ups of small herbs or animals. Future work may automate the search using image recognition, something Instagram’s terms of use now permit. Most importantly, the combination of social media validation and satellite classification yields not just presence-only points but presence-absence data across the entire latitudinal range of a treeline species, precisely the foundation needed for modeling how lenga’s upper limit will shift as the climate warms. For a species whose radial growth and seedling establishment respond sensitively to climate variation, and for the countless remote mountain species that tourists photograph but scientists rarely reach, the message is clear: the crowdsourced camera roll has become a legitimate scientific instrument.

Subject of Research: Using Instagram posts as ground truthing data to validate remote sensing maps of Nothofagus pumilio distribution in the southern Andes

Article Title: Instagram data for validating Nothofagus pumilio distribution mapping in the Southern Andes: A novel ground truthing approach

Article References: Werner, M., Weidinger, J., Böhner, J., Schickhoff, U., & Bobrowski, M. (2024). Instagram data for validating Nothofagus pumilio distribution mapping in the Southern Andes: A novel ground truthing approach. Frontiers of Biogeography, 17, Article e140606. https://doi.org/10.21425/fob.17.140606

Image Credits: AI Generated

DOI: 10.21425/fob.17.140606

Keywords: Instagram, Nothofagus pumilio, treeline, remote sensing, Sentinel-2, citizen science, GBIF, species distribution mapping, Patagonia, supervised classification, sampling bias, biogeography