This opinion article was written by GWSC Environmental Data Scientist Dr. Sambadi Majumder.
In recent years, climate researchers have increasingly leveraged Machine Learning (ML) models to answer key research questions. ML models are generally adept at learning from complex weather and climate data and provide reliable forecasts.
However, these models are also characterized by their general opaqueness to the learning process in relation to the data. This “black box” nature results in superior predictive performance when compared to traditional statistical models but knowledge about scientifically relevant patterns within the data remains elusive (Molnar et al., 2020). This lack of transparency into the model’s reasoning can result in a lack of trust by the end-user and impact their general adoption.
In climate and environmental sciences, interpretability is not just beneficial, it is imperative. It enables scientists and policymakers to unravel complex climate systems, validate predictions, and foster stakeholder trust, thereby facilitating informed decision-making in addressing climate variability and change.
“Explainable AI” (XAI) or “Interpretable ML” (IML) is a branch of ML research largely dedicated to formulating frameworks that facilitate interpretability within ML models. From studying the dynamic environmental processes inextricably linked to weather and air quality to providing actionable insights for better water quality management, researchers have implemented XAI in a myriad of ways.
Examples of this include developing and implementing interpretable “Deep Learning” models that are specialized to elucidate the influence of meteorological factors and regional characteristics on particulate matter concentrations, offering a window into the environmental dynamics affecting air quality (Yan et al., 2021; Gu et al., 2021). This approach of using explainable “Deep Learning” models has also been used in other areas of climatology, specifically in understanding intricate climate patterns and enabling researchers to make meaningful geoscientific discoveries (Toms et al., 2020).
Additionally, traditional machine learning algorithms tweaked to be more interpretable have been used alongside such interpretable “Deep Learning” models to extrapolate relevant environmental patterns that impact precipitation in the Western US (Gibson et al., 2021). In response to urban development, researchers at the University of Michigan and Michigan State, have been able to highlight the importance of high-density development on water pollution reduction in the Texas Gulf region, by using XAI ML techniques (Wang et al. 2021). Another interesting application of IML within the sphere of environmental sciences includes identifying important water and climate related factors that contribute to evapotranspiration in crops (Chakraborty et al., 2021).
The integration of IML into climate and environmental sciences has the potential to revolutionize our approach to understanding and managing the Earth’s complex systems. By breaking the “black box” of traditional ML models, researchers are not only enhancing the model results but also building the necessary trust and transparency for these technologies to be widely accepted and applied.
Through innovative applications across various environmental challenges, this field is paving the way for more informed, data-driven decisions critical to addressing the pressing issues of climate change and environmental sustainability. This journey towards interpretability in ML signifies a vital step forward in harnessing the full potential of artificial intelligence for the greater good of our planet and future generations.