Exploring machine learning potential for climate change risk assessment

Abstract

Global warming is exacerbating weather, and climate extremes events and is projected to aggravate multi-sectorial risks. A multiplicity of climate hazards will be involved, triggering cumulative and interactive impacts on a variety of natural and human systems.An improved understanding of risk interactions and dynamics is required to support decision makers in their ability to better manage current and future climate change risks. To face this issue, the research community has been starting to test new methodological approaches and tools, including the application of Machine Learning (ML) leveraging the potential of the large availability and variety of spatio-temporal big data for environmental applications. Given the increasing attention on the application of ML methods to Climate Change Risk Assessment (CCRA), this review mapped out the state of art and potential of these methods to this field of research. Scientometric and systematic analysis were jointly applied providing an in-depth review of publications across the 2000–2020 timeframe. The resulting output from the analysis showed that a huge variety of ML algorithms have been already applied within CCRA, among them, the most recurrent are Decision Tree, Random Forest, and Artificial Neural Network. These algorithms are often applied in an ensemble or hybridized way to analyze most of all floods and landslides risk events. Moreover, the application of ML to deal with remote sensing data is consistent and effective across reviewed CCRA applications, allowing the identification and classification of targets and the detection of environmental and structural features. On the contrary concerning future climate change scenarios, literature seems not to be very widespread into scientific production, compared to studies evaluating risks under current conditions. The same lack can be noted also for the assessment of cascading and compound hazards and risks, since these concepts are recently emerging in CCRA literature but not yet in combination with ML-based applications.