Normally when one hears the term ‘artificial intelligence’, or deep neural networks, one might think of the technology giants and their ability to crunch immense amounts of data through our personal browsing habits so that they can build profiles of us in order to shill more content and product our way.
But artificial intelligence (AI) is a multi-faceted hydra, one which has permeated many aspects of our lives and the technologies we interact with and one which the broader research community is engaging with in new and novel ways.
A good example of this is the recently published paper supported by the CHARTER project lead authored by Rasmus Erlandsson which utilized AI to estimate in a new and important way, lichen volume, something that was difficult to quantify with more traditional image interpretation methods. Using AI and machine learning to quantify biomass is a very new endeavor, and this publication is arguably the first to do so in a systematic manner.
Erlandsson and his colleagues trained neural networks with Landsat 5, 7 and 8 images (insert link) and ground-truthed data collected over 20 years in Norway, Sweden, Finland and the Kola peninsula. While there are higher resolution images now available, the extended time series available via Landsat imagery makes a multi decadal overview (back to the 1980’s) of change more possible, and also revealing in terms of the possible applications of this approach. Erlandsson and team utilized over this geographical space a total of 8914 points of data, which given the immense amount of data that current algorithms crunch, is a tiny slice of data, albeit in a rather niche field, and work demanding to collect.
Like all new algorithms, the model needed ‘training’, and embedded in the article is the rather nice phrase ‘blunt naivety’. In other words, by using an AI approach – the challenges of creating a workable algorithm notwithstanding, there are benefits because one does away with prior assumptions regarding data interpretation of lichen reflectance in satellite imagery, and in doing so, create an enhanced ability to find context dependent relationships – such as landscape and gradient shifts across open tundra and boreal landscapes. AI offers the possibility to crunch multiple data sources across time and space far more efficiently in terms of time than a person would.
Erlandsson noted that he was fortunate to have a sibling who works in the field of AI and they communicated mostly by phone during the pandemic, but this did not prevent the creation of this rather complex but intriguing graphic, which represents the creation of a neural network made graphic and the steps that the network takes before positing conclusions.
Erlandsson stated that the success of this approach holds out the potential for a broader approach, with more data inputs and a broader geographical scope. The paper is very new so wider feedback is expected and anticipated. As the paper notes the model developed had some challenges in distinguishing sand from lichen, but Erlandsson noted in a follow up conversation that they have already trained a model that is better at distinguishing the two. Another issue is that it is quite time consuming to assess large areas and Erlandsson hopes to speed up the work flow and processing time in follow up studies.
Read the paper here, visit Rasmus’ personal webpage here. He is a postdoctoral researcher in Ecology at the Norwegian Institute for Nature Research (NINA) and at the University of Tromsø, Norway and his research is focused mainly on landscape ecology in subarctic and arctic ecosystems, based on – or related to – remote sensing. He was a visiting reasearcher at the Environmental Change Research Unit (ECRU) at Helsinki University in Finland from August 2021 to July 2022.