Explainable Artificial Intelligence and Climate Change

Posted on by David Parkes

world map with environmental information

What are the applications of machine learning (used here interchangeably with 'AI') in understanding global climate change? Why do machine learning algorithms reach the conclusions they do, and when do they get things wrong? What are the limits of what machine learning can tell us, and where do we need to involve human environmental science expertise? These were the big questions addressed by Elizabeth (Libby) Barnes, of Colorado State University, in her seminar delivered to DSNE on Thursday 30th September. In this post I want to cover the key ideas embodied in Libby's work, and provide commentary – some Libby's, some my own – on how they fit into the landscape of environmental science.

At the core of the work she presented, Libby introduced the idea of explainable artificial intelligence (XAI); methods to take the results from machine learning algorithms and trace the path back to the input data, showing which elements of the data were most important in decision-making. These allow a critical assessment of the AI's performance: do the components of the dataset that lead to the results match our human, process-based understanding of what drives them? If not, the AI may reach the right conclusion for the wrong reason. Sometimes the AI getting it right or wrong is clear – in image processing, for example, a heat map of which parts of an image are most relevant in identifying the presence of an animal immediately shows whether the AI has correctly identified the animal-containing sections of the image – but in many cases with environmental data, interpretation is more difficult. This is where the skills of subject matter experts are indispensable: squaring the AI-generated outputs with process-based understanding is critical to trusting the meaning-agnostic results of an algorithm, while the insights offered by XAI can give scientists a way to approach datasets that are so large, complex, or noisy that it's hard to know where to start in offering a full interpretation.

Armed not only with the AI tools to process large-scale data, but also with XAI to understand where the results have come from, Libby took us through examples where she and her team had applied machine learning to interpret major global changes in the earth system. The first example was of the Human Footprint Index (HFI), a measure of human pressure on the environment. To provide a comprehensive perspective, HFI incorporates many different datasets – land use, population density, light pollution, etc. – and as a result it is compiled infrequently with a significant lag behind the present day (at least in the context of rapid modern changes in human environmental impact). Libby and her colleagues trained an AI to generate estimates of HFI based on Landsat imagery – a single, more straightforward dataset – and used this to estimate a more up-to-date HFI for 2019 (Keys, Barnes, and Carter, 2021). But is it any good? This is where XAI comes in. Using the technique of Layerwise Relevance Propagation (LRP), heat-maps of the parts of the dataset that were most important to the AI in determining changes between the 2000 HFI and the 2019 estimate show the AI recognising the presence of many new man-made features – wind farm installations in Texas, irrigation networks in Cairo, urban development in Hanzhong – in the more recent Landsat images. These are tiny examples within a global dataset, of course – to cast a human eye over every pixel would defeat the point of the time-saving and scalability benefits that are primary motivators for using AI at all – but they are critical in building trust in the AI output.

The second example presented was of an AI trained on gridded global datasets of temperature to predict the year (Barnes et al. 2020). A strange application at first glance – most of us are aware of what year it is, and of what year any data we record is from – but a useful case study for using XAI to interpret results in a fairly controlled (at least in earth sciences terms) environment. The initial results are as expected: the AI struggles to predict the year before the 1960s or 70s, when the global temperature trend was smaller in comparison to noise, and then makes effective predictions once the signal of forced temperature change is clearer. Things become more interesting when XAI is used to look at why the AI predicts the year it does. The key areas for predicting the year are not necessarily those where annual temperature anomaly is of the highest amplitude, but where the signal-to-noise ratio is particularly high. Is this, then, significantly different from a more simple regression between years and relative temperature anomalies? For me the answer is both yes and no. The AI-generated results are perhaps less of a brand-new perspective and more an extension of what is possible with more standard statistical techniques, but the ability to consider the data as a whole rather than point-by-point confers the ability for the AI to recognise spatial patterns in a way that – without an explicitly applied framework that then requires additional assumptions – the simpler methods cannot. The amount of added value that the AI provides here relies, it seems, on the internal details of how the algorithm used identifies patterns in data; a matter of considerable interest, but better left to the paper rather than covered in either a presentation or a blog post.

In the third example presented, we saw machine learning used to feed back more explicitly into our process-based understanding of climate change. Libby demonstrated work by a member of her team, Emily Gordon, on using AI to make predictions of transition in the Pacific Decadal Observation (PDO) (Gordon, Barnes, and Hurrell, 2021). Here, the nature of what XAI is telling us is less immediately intuitive, and for good reason: a global dataset (of ocean heat content) is condensed down to just a single value representing likelihood of transition. Nevertheless, the results are compelling: through XAI there is a coherent picture of the areas which are most relevant to both directions of transition for the PDO, and they are in agreement with current theory.

The use of AI as a way of directing attention in predicting and understanding complex emergent behaviours in large-scale climate variability, rather than as simply a way to process and characterise data, does have a feel (at least to a relative newcomer such as myself) of being more in its infancy, but it is also potentially the most exciting application. One of the most profound challenges of climate science is dealing with the vast number of interacting processes and systems, which are extremely hard to isolate and often non-linear, unintuitive or even chaotic in their responses. Human scientists can specialise with more depth in individual subject areas, or choose to focus on broader but shallower knowledge across multiple disciplines, but ultimately the complexity of many phenomena may outstrip the capacity for any individual scientist or group to reasonably understand all of the relevant causal relationships involved. How, then, do we start explaining things 'too big' to explain? There is a relevant analogy here from the world of pure mathematics in the form of the Four Colour Theorem, which was the first major case of a proof being performed by a computer which could not be feasibly checked directly by human mathematicians because it was so large. It was resisted by some, but the proof still stands, with the verification of the proof-generating software standing in for the ability to verify the proof itself. Is environmental science in a position to do similarly when it comes to understanding complex, large-scale patterns; to verify the coding of the result-generating AI rather than of a directly-stated physical explanation? Probably not – or not yet – but we are likely to see steps in that general direction, with AI at the very least starting to direct where subject-matter experts begin when explaining increasingly complex phenomena.

The potential for speculation on future AI uses in environmental science is almost endless – in direct contrast to the length of a good blog entry – but it is clear from the work already performed and showcased by Libby that this is a fantastically productive area, not only helping us to handle data but also, thanks to the insights of XAI, helping us to explain it. On behalf of DSNE I'd like to thank Libby for her time and energy, and for getting us all thinking about both the power and the pitfalls of using AI in environmental science.

Elizabeth Barnes is an associate professor at Colorado State University. Information on her research group is available at https://sites.google.com/view/barnesgroup-csu/

Header image credit: Keys, Barnes, and Carter (2021)

References:

Keys, Barnes, and Carter (2021), Environmental Research Letters

https://iopscience.iop.org/article/10.1088/1748-9326/abe00a

Barnes et al. (2021), Journal of Advances in Modeling Earth Systems

https://doi.org/10.1029/2020MS002195

Gordon, Barnes, and Hurrell (2021), Geophysical Research Letters (in review) https://doi.org/10.1002/essoar.10507680.2

Related Blogs

Disclaimer

The opinions expressed by our bloggers and those providing comments are personal, and may not necessarily reflect the opinions of Lancaster University. Responsibility for the accuracy of any of the information contained within blog posts belongs to the blogger.

Back to blog listing