Accurate estimation of drought events is vital for the mitigation of their adverse consequences on water resources, agriculture and ecosystems. Machine learning algorithms are promising methods for drought prediction as they require less time, minimal inputs, and are relatively less complex than dynamic or physical models. In this study, a combination of machine learning with the Standardized Precipitation Evapotranspiration Index (SPEI) is proposed for analysis of drought within a representative case study in the Tibetan Plateau, China, for the period of 1980-2019. Two timescales of 3 months (SPEI-3) and 6 months (SPEI-6) aggregation were considered. Four machine learning models of Random Forest (RF), the Extreme Gradient Boost (XGB), the Convolutional neural network (CNN) and the Long-term short memory (LSTM) were developed for the estimation of the SPEIs. Seven scenarios of various combinations of climate variables as input were adopted to build the models. The best models were XGB with scenario 5 (precipitation, average temperature, minimum temperature, maximum temperature, wind speed and relative humidity) and RF with scenario 6 (precipitation, average temperature, minimum temperature, maximum temperature, wind speed, relative humidity and sunshine) for estimating SPEI-3. LSTM with scenario 4 (precipitation, average temperature, minimum temperature, maximum temperature, wind speed) was relatively better for SPEI-6 estimation. The best model for SPEI-6 was XGB with scenario 5 and RF with scenario 7 (all input climate variables, i.e., scenario 6 + solar radiation). Based on the NSE index, the performances of XGB and RF models are classified as good fits for scenarios 4 to 7 for both timescales. The developed models produced satisfactory results and they could be used as a rapid tool for decision making by water-managers.

State-of-the-art climate models project a substantial decline in precipitation for the Mediterranean region in the future¹. Supporting this notion, several studies based on observed precipitation data spanning recent decades have suggested a decrease in Mediterranean precipitation^2,3,4, with some attributing a large fraction of this change to anthropogenic influences^3,5. Conversely, certain researchers have underlined that Mediterranean precipitation exhibits considerable spatiotemporal variability driven by atmospheric circulation patterns^6,7 maintaining stationarity over the long term^8,9. These conflicting perspectives underscore the need for a comprehensive assessment of precipitation changes in this region, given the profound social, economic and environmental implications. Here we show that Mediterranean precipitation has largely remained stationary from 1871 to 2020, albeit with significant multi-decadal and interannual variability. This conclusion is based on the most comprehensive dataset available for the region, encompassing over 23,000 stations across 27 countries. While trends can be identified for some periods and subregions, our findings attribute these trends primarily to atmospheric dynamics, which would be mostly linked to internal variability. Furthermore, our assessment reconciles the observed precipitation trends with Coupled Model Intercomparison Project Phase 6 model simulations, neither of which indicate a prevailing past precipitation trend in the region. The implications of our results extend to environmental, agricultural and water resources planning in one of the world’s prominent climate change hotspots¹⁰.

This study examined the physical properties of agricultural drought (i.e., intensity, duration, and severity) in Hungary from 1961 to 2010 based on the Standardized Precipitation Index (SPI) and the Standardized Precipitation Evapotranspiration Index (SPEI). The study analyzed the interaction between drought and crop yield for maize and wheat using standardized yield residual series (SYRS), and the crop-drought resilient factor (CDRF). The results of both SPI and SPEI (-3, -6) showed that the western part of Hungary has significantly more prone to agricultural drought than the eastern part of the country. Drought frequency analysis reveals that the eastern, northern, and central parts of Hungary were the most affected regions. Drought analysis also showed that drought was particularly severe in Hungary during 1970–1973, 1990–1995, 2000–2003, and 2007. The yield of maize was more adversely affected than wheat especially in the western and southern regions of Hungary (1961–2010). In general, maize and wheat yields were severely non-resilient (CDRF < 0.8) in the central and western part of the country. The results suggest that drought events are a threat to the attainment of the second Sustainable Development Goals (SDG-2). Therefore, to ensure food security in Hungary and in other parts of the world, drought resistant crop varieties need to be developed to mitigate the adverse effects of climate change on agricultural production.

Drought is a fundamental physical feature of the climate pattern worldwide. Over the past few decades, a natural disaster has accelerated its occurrence, which has significantly impacted agricultural systems, economies, environments, water resources, and supplies. Therefore, it is essential to develop new techniques that enable comprehensive determination and observations of droughts over large areas with satisfactory spatial and temporal resolution. This study modeled a new drought index called the Combined Terrestrial Evapotranspiration Index (CTEI), developed in the Ganga river basin. For this, five Machine Learning (ML) techniques, derived from artificial intelligence theories, were applied: the Support Vector Machine (SVM) algorithm, decision trees, Matern 5/2 Gaussian process regression, boosted trees, and bagged trees. These techniques were driven by twelve different models generated from input combinations of satellite data and hydrometeorological parameters. The results indicated that the eighth model performed best and was superior among all the models, with the SVM algorithm resulting in an R² value of 0.82 and the lowest errors in terms of the Root Mean Squared Error (RMSE) (0.33) and Mean Absolute Error (MAE) (0.20), followed by the Matern 5/2 Gaussian model with an R² value of 0.75 and RMSE and MAE of 0.39 and 0.21 mm/day, respectively. Moreover, among all the five methods, the SVM and Matern 5/2 Gaussian methods were the best-performing ML algorithms in our study of CTEI predictions for the Ganga basin.