Pampas grass, a feathery ornamental plant native to South America, has been spreading far beyond gardens, invading ecosystems across the globe. In Portugal, this tenacious grass has steadily expanded its range, displacing native plants and disrupting habitats.

Keeping tabs on the relentless march of invasive plants like pampas grass is critical for protecting biodiversity, but it’s easier said than done. Traditional monitoring methods, which rely on boots-on-the-ground surveys and experts painstakingly identifying plants, are costly and time-consuming.

But what if we could crowdsource this Herculean task by tapping into the millions of plant photos shared on social media every day? Ana Sofia Cardoso and colleagues have tried harnessing the power of artificial intelligence to scour social networks for images of the offending grass.

Their AI plant detectives, trained on expert-identified photos from citizen science databases, proved remarkably adept at picking out pampas grass in all kinds of images. The results, published in the journal Ecological Informatics, hint at a new paradigm for monitoring the spread of invasive species: one that’s faster, cheaper, and more scalable than ever before.

Meet the AI plant detectives: a trio of deep learning models with a keen eye for pampas grass.

To train their algorithms to accurately identify this invasive plant, the research team started by feeding them expertly-labeled photos from citizen science platforms like Invasoras.pt and iNaturalist. These images, painstakingly annotated by knowledgeable volunteers, provided a gold standard for what pampas grass looks like in the wild.

The team put three different deep learning architectures through their paces: DenseNet201, Faster R-CNN ResNet50, and Faster R-CNN Inception-v2. The first model, a classification specialist, learned to label images as either containing pampas grass or not. The other two, object detection models, went a step further by learning to draw bounding boxes around the plant in images.

After extensive training, the models were put to the test on a new set of citizen science images. The results were impressive: the best-performing models, DenseNet201 and Faster R-CNN ResNet50, correctly identified pampas grass more than 94% of the time. When the algorithms made mistakes, it was often on trickier images where the grass was small, blurry, or in the background.

But the real challenge was yet to come. The researchers wanted to see if their AI plant detectives could spot pampas grass “in the wild” – not in curated citizen science photos, but in the unfiltered stream of images posted to social media. They set the algorithms loose on hundreds of images scraped from Instagram, Flickr, Twitter, and Facebook.

Remarkably, the models held their own, correctly identifying pampas grass in more than three-quarters of the social media photos. The performance dip, the researchers believe, stems from the lower quality and resolution of many social media images. These results suggest that deep learning models, trained on a relatively small set of photos, can effectively translate those learnings to the unstructured world of social media imagery.

Why Instagram?

Sites like Instagram can provide photos in volume, but a photo alone is not enough. What makes a photo shared on social media so useful is the data that comes with it. Many images shared on platforms like Instagram and Flickr come with embedded location data, latitude and longitude coordinates that pinpoint exactly where the photo was taken. By extracting these geotags from images flagged as containing pampas grass, the researchers could chart the invasive plant’s distribution with unprecedented precision.

The team focused their efforts on photos posted between 2019 and 2021, a time of rapid expansion for pampas grass in Portugal. As they’d hoped, the AI-generated maps revealed a number of previously undocumented pampas grass sightings, especially in the country’s north, coastline, and southern regions.

Spatial distribution of Cortaderia selloana: (a) for Flickr across 2019, 2020 and 2021, (b) for Instagram, Flickr and Invasoras.pt. Black circles and red arrows indicate new potential locations in relation to the data available on Invasoras.pt. Cardoso et al. 2024

Comparing the geotagged detections year-over-year painted a troubling picture: pampas grass was on the move, popping up in new locations and filling in its range with each passing year. From just a scattering of detections in 2019, the grass had spread to more than double the number of locations by 2021.

While the findings are worrying from a conservation perspective, they showcase the potential of this AI-powered, social media-driven approach to invasive species monitoring. By revealing pampas grass hotspots and tracking the plant’s spread in near-real time, these maps could help guide critical early detection and rapid response efforts.

AI species screening is not something like Conservation-GPT

AI is very much the 2024 buzzword, but the AI used by Cardoso and colleagues is very different to the large language models, like Chat-GPT that are grabbing headlines. Cardoso and colleagues emphasise the importance of human intelligence in the system. The pampas grass photos used to train the AI are explicitly identified using the expertise of the iNaturalist community. The researchers also highlight some limitations.

For one, the AI models are currently only reliable at identifying pampas grass when it’s in full feathery bloom. During other stages of its life cycle, when the plant lacks its distinctive plumes, it can be tough to distinguish from other grass species. This means that some pampas grass populations, especially younger or recently-established ones, might fly under the AI’s radar.

Humans also limit where the data comes from. People tend to take and share photos in certain areas, like cities, parks, and tourist attractions, more than others. This means that the AI-generated maps might over-represent pampas grass populations in these popular spots, while missing sightings in more remote or less-frequented areas.

To address these challenges, the researchers are already hard at work on version 2.0 of their AI plant detectives. They’re exploring ways to train the models on a more diverse set of pampas grass images, showing the plant at different life stages and in a wider variety of habitats. They’re also looking into methods to account for and correct the inherent biases in geotagged social media data.

The researchers stress that their AI tools are meant to complement, not replace, the expertise of human ecologists and conservationists. But by automating certain tedious or time-consuming tasks, like sifting through thousands of photos, these algorithms could free up valuable human resources to focus on higher-level strategy and on-the-ground action.

There’s also the possibility of extending this approach to other invasive species, from algae to zebra mussels. While the specifics would differ, the core idea – using social media data and AI to map invasions in real-time – could be a game-changer for the field of invasion ecology.

Researchers caution there are ethical considerations

These applications have the potential to support the identification of priority areas for eradication efforts, the efficient allocation of resources, and the evaluation of the success of management interventions over time. Still, we are also aware of potential barriers in the acceptability and confidence of using artificial intelligence tools and user-generated contents by these organizations, especially in the context of social issues like ethics and fairness.

Cardoso et al. 2024

The ethical question is whether it is acceptable for scientists to use public social media images. The images on iNaturalist are submitted with the intention of helping scientific research. Images on Instagram or other social media sites are uploaded for many reasons. The power of using these sites is the sheer number of images that can be scanned. However, this volume of data also means that seeking active consent for all images is impractical. Can we assume that people don’t object to their photos being used to track invasive species?

The poison in this assumption is there are other tools that use AI photo analysis for facial recognition. One such site happily links to news stories about how its tools can be used for cyberstalking. It’s easy to see why there might be a lack of trust in such systems. For this reason, Cardoso and colleagues state: “it is essential to highlight the transparency and fairness in the overall workflow adopted, addressing any biases or ethical concerns associated with deep learning applications and the use of personal data.”

If these ethical concerns can be addressed, then photo-scanning could prove to be a valuable tool for fighting invasive species. If now

READ THE ARTICLE:

Cardoso, A.S., Malta-Pinto, E., Tabik, S., August, T., Roy, H.E., Correia, R., Vicente, J.R. and Vaz, A.S. (2024) “Can citizen science and social media images support the detection of new invasion sites? A deep learning test case with Cortaderia selloana,” Ecological Informatics, (102602), p. 102602. Available at: https://doi.org/10.1016/j.ecoinf.2024.102602.


Cover image: Pampas grass by JLPC / Wikimedia Commons.

Laisser un commentaire

Votre adresse e-mail ne sera pas publiée. Les champs obligatoires sont indiqués avec *