Advancing the detection of multi-sector drought impacts via feature extraction and multi-task learning

Drought indices are essential tools for quantifying drought conditions by integrating multiple variables into a single measure that represents its characteristics, such as intensity, duration, and severity. These indices play a key role in real-time monitoring, forecasting, and supporting risk management actions. However, traditional statistical indices often fail to account for the complex interactions between drought precursors and their socio-economic and environmental impacts. Moreover, given the absence of a universally accepted drought definition, no single index is applicable to all drought types, climate conditions, or affected sectors.

In this study, we aim to improve traditional drought detection by defining new impact-based drought indices through Machine Learning algorithms. These indices are designed to better link the observed impacts of extreme droughts across different sectors with their potential drivers, including climatic, meteorological, and hydrological variables, analyzed across multiple spatial and temporal scales. The methodology is applied to the case study of the Adda River basin, focusing on the multisectoral impacts of drought on agricultural production, hydroelectric generation, and recreational and ecosystem services.

The definition of impact-based drought indices relies on the FRamework for Index-based Drought Analysis (FRIDA), which uses a feature extraction algorithm to formulate novel impact-based drought indices that combine all the relevant information about candidate drought drivers (e.g. water levels, snow depth, temperature) to reproduce the observed impacts.

Our findings indicate that FRIDA has produced indices that accurately capture the drought impacts with the Pearson correlation coefficient between observations and model’s outputs that remains consistently above 0.6, with values reaching 0.97 and 0.99 for the hydropower and recreation sectors, respectively. Additionally, it is noteworthy that the inputs selected by the algorithm vary depending on the sector being considered, shedding light on sector-specific connections between drivers and impacts. Ongoing experiments are investigating the potential for further improving our results by adopting a multi-task model for better handling the interdependencies across the impacted sectors with respect single-task models that identify individual indices independently for the different sectors.