Heatwaves can greatly impact societies, underscoring the need to extend current heatwave prediction lead times. This study investigates multiple machine learning (ML) model approaches for heatwave occurrence prediction with long lead times of one to five months. Five ML classifiers, built using Google Earth Engine remote sensing datasets, are developed and tested for heatwave prediction for the national scale (case example of Sweden) over time period 1989–2019. The ML modelling is based on 13 final explanatory atmospheric and landscape features. The balanced random forest model exhibits the consistently best performance among the tested ML models, stable across all investigated lead times (from one to five months) with balanced accuracy of around 0.77, even though not overall identifying actual heatwave occurrence (decreased recall for heatwave occurrence from 0.87 to 0.81). Application of SHapley Additive exPlanations technique for model interpretation shows increasing importance of model output with increasing lead time for landscape features such as runoff and soil water. Overall, more frequent heatwave occurrence emerges for places characterized by lower values of geopotential height, evaporation, precipitation, and topographical slope, and higher values of temperature, runoff, and sea level pressure. The study also exemplifies how the developed ML modelling approach could be used to identify and warn for early signs of forthcoming heatwave occurrence, and further step-wise improve the identification and warning toward less uncertainty for shorter lead times. This can facilitate earlier warning in support of better planning of measures to mitigate adverse heatwave impacts, up to several months ahead of their possible occurrence.

In this paper we present a framework to aid in the selection of optimal environmental indicators for detecting and mapping extreme events and analyzing trends in heatwaves, meteorological and hydrological droughts, floods, and their compound occurrence. The framework uses temperature, precipitation, river discharge, and derived climate indices to characterize the spatial distribution of hazard intensity, frequency, duration, co-occurrence, and dependence. The relevant climate indices applied are Standardized Precipitation Index, Standardized Precipitation and Evapotranspiration Index (SPEI), Standardized Streamflow Index, heatwave indices based on fixed (HWI $$_\textrm{S}$$ S ) and anomalous temperatures (HWI $$_\textrm{E}$$ E ), and Daily Flood Index (DFI). We selected suitable environmental indicators and corresponding thresholds for each hazard based on estimated extreme event detection performance using receiver operating characteristics (ROC), area under curve (AUC), and accuracy, which is defined as the proportion of correct detections. We assessed compound hazard dependence using a Likelihood Multiplication Factor (LMF). We tested the framework for the case of Sweden, using daily data for the period 1922–2021. The ROC results showed that HWI $$_\textrm{S}$$ S , SPEI12 and DFI are suitable indices for representing heatwaves, droughts, and floods, respectively (AUC > 0.83). Application of these indices revealed increasing heatwave and flood occurrence in large areas of Sweden, but no significant change trend for droughts. Hotspots with LMF > 1, mostly concentrated in Northern Sweden from June to August, indicated that compound drought-heatwave and drought-flood events are positively correlated in those areas, which can exacerbate their impacts. The novel framework presented here adds to existing hydroclimatic hazard research by (1) using local data and historical records of extremes to validate indicator-based hazard hotspots, (2) evaluating compound hazards at regional scale, (3) being transferable and streamlined, (4) attaining satisfactory performance for indicator-based hazard detection as demonstrated by the ROC method, and (5) being generalizable to various hazard types.

Agricultural drought can severely reduce crop yields, lead to large economic losses and health impacts. Combined climate and land use variations determine key indicators of agricultural drought, including soil moisture and the Palmer drought severity index (PDSI). This study investigated the use of machine learning (ML) methods for predicting these indicators over Sweden, spanning steep climate and land use gradients. Three data arrangement methods (multi-features, temporal, and spatial) were used and compared in combination with seven ML/deep learning (DL) models (random forest (RF), decision tree, multivariate linear regression, support vector regression, autoregressive integrated moving average (AMIRA), artificial neural network, and convolutional neural network). Seven investigated features, obtained from Google Earth Engine, were used in the ML/DL modeling (soil moisture, PDSI, precipitation, evapotranspiration, elevation, slope and soil texture). The temporal ARIMA model (found most suitable for local scale prediction) and the multi-features RF model (more suitable for national-scale prediction) emerged as best performing for soil moisture prediction (with MAE of 9.1 and 11.95, and R2 of 0.79 and 0.59, respectively). All models generally performed better in predicting the soil moisture than the PDSI indicator of drought. For drought indicator prediction and mapping, previous-year average monthly soil moisture emerged as the most important feature, combined with the four additional corresponding features of PDSI, precipitation, evapotranspiration and elevation.