Understanding impacts of compound drought and heatwaves: A multi-risk analysis on agricultural-dominated socio-ecological systems combining Earth Observation and Machine Learning

Challenges from cascading and compounding multi-hazard events are increasing, reinforcing the need to integrate emerging technologies into current risk assessment methodologies to disentangle the complexity of multi-risk events and support adaptation strategies. This work explored the integration of Earth Observation (EO) and machine learning to advance our understanding of multi-risk drought and heatwave events, with a case study in the downstream Adige River basin (Northern Italy). The study aimed at understanding the root causes of multi-risk drought and heatwave impacts and their spatiotemporal dynamics. To do so, a novel multi-risk assessment approach was adapted from the Forensic Investigation of Disasters (FORIN) framework and focused on the investigation of the upstream-downstream impact dynamics, with a testbed on irrigated agriculture. Two summer cropping periods with contrasting drought conditions (2022 and 2023, more and less dry respectively) were analysed, allowing for a spatiotemporal comparative investigation that sought to understand how differential traits (e.g., hazards, vulnerabilities) in similar settings (e.g., same area, different time) could explain the observed impacts. First, hydrometeorological hazards (SPEI 90-, 180-, and 365-days, temperature anomalies, river discharge, EO based soil moisture), exposure (spatiotemporal mapping of maize presence with in situ and EO data), and vegetation conditions were characterized. The latter were considered as a proxy of impact and calculated using Sentinel-2 derived indices (NDVI, NDMI) combined with a Principal Component Analysis into a composite stress index, then aggregated at the crop-field level (~20,000 fields in total) for each satellite image. Subsequently, unsupervised machine learning (HDBSCAN – Hierarchical Density-Based Spatial Clustering of Applications with Noise) was applied on the composite stress index for each date to identify field clusters and homogeneous areas having consistent vegetation conditions. Finally, clusters were further categorized into impacted or not impacted based on empirical stress thresholds rooted in NDVI and NDMI, to produce a final susceptibility map that represented the spatial frequency of stress occurrence. Results revealed a clear upstream–downstream stress gradient along the river, well summarized by the susceptibility map. This trend was mostly evident and statistically significant in 2022, and proved in line with upstream-downstream river discharge differences and provincial level yield data. Given the comparable hazard conditions along the case study, this suggested that additional factors might have had a strong influence on driving impacts, such as irrigation water management along the river. Additionally, correlation analysis revealed weak relationships between the composite stress index and the other variables (e.g., hazards, soil properties, field position) within the same year, suggesting the presence of complex socio-ecological aspects and physical vulnerabilities (e.g., water management, groundwater availability) that shaped vegetation stress beyond the hazard itself. Considering the large amount of data and the high resolution and scale of the analysis, this study advances the understanding of spatiotemporal dynamics of multi-risk events through spatiotemporal representation of impact dynamics, highlighting the added value of integrating EO and in situ data with machine learning techniques to unravel underlying vulnerability factors and enhance multi-hazard risk assessment to support adaptation and management strategies.