91��ý

2012. 2018. Farmers in Northern Kazakhstan remember those years well. In some districts, crop losses reached 40%. The fields stood dry without rain, grain failed to mature, soil moisture disappeared—and there was no warning.

This vulnerability to sudden drought became the starting point for a new scientific project. "Assessment of Atmospheric Drought Risk and Development of an Early Warning System for Northern Kazakhstan Based on Machine Learning" is being carried out during 2025–2027 within the priority area "Ecology, Environment and Sustainable Natural Resource Management." The project is a fundamental research study in the field of Earth Sciences.

Much of Kazakhstan lies within arid and semi-arid climatic zones. Northern Kazakhstan is the country's principal grain-producing region. In 2020, Kazakhstan ranked fifth in the world in wheat exports, shipping 7.4 million metric tonnes. During 2022–2023, national wheat production reached 16.4 million tonnes, placing the country ninth among the world's largest producers. However, agriculture in the region depends almost entirely on natural rainfall. Irrigation infrastructure remains limited, meaning that drought directly threatens crop production.

Global warming is making this dependence increasingly dangerous. According to reports by the Intergovernmental Panel on Climate Change (IPCC) and numerous international studies, climate change is increasing both the frequency and intensity of extreme weather events, including droughts. Over recent decades, Central Asia has experienced a significant rise in average annual temperatures accompanied by only a slight decline in precipitation. Approximately 75% of the region is considered highly vulnerable to natural disasters. Worldwide estimates indicate that up to 80% of agricultural economic losses in both developing and developed countries are caused by drought, while the global area affected by drought has doubled over the past forty years.

Kazakhstani researchers have previously investigated this issue. Using meteorological records from 1971 to 2008, they analysed the frequency of atmospheric droughts in Northern Kazakhstan employing several drought indices, including Selyaninov's Hydrothermal Coefficient (HTC), SPI, Shashko's Index, and Ped's Index. A long-term drought forecasting methodology was subsequently developed.

However, these earlier studies had important limitations. First, traditional forecasting approaches—including analogue methods and numerical climate models—achieved prediction accuracies of only 50–70%, particularly performing poorly for precipitation forecasts. Second, the analyses were conducted only at the regional scale, without considering climatic differences between individual districts.

Yet climatic conditions can vary considerably even within the same region. A forecast produced for "Northern Kazakhstan as a whole" cannot provide reliable guidance for a specific district. Addressing this gap is the project's primary objective.

The project differs fundamentally from previous research in three major ways.

The first innovation is its algorithms. Instead of relying solely on traditional statistical methods, the researchers employ advanced machine learning techniques, including GRU (Gated Recurrent Unit) neural networks for meteorological time-series forecasting and Random Forest and XGBoost algorithms for classifying drought and non-drought conditions. These models are capable of identifying complex nonlinear relationships between climatic variables and drought occurrence that conventional statistical approaches cannot detect.

The second innovation lies in its data sources. Alongside meteorological observations, the project integrates satellite remote sensing data, including the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI) obtained from the Copernicus and EUMETSAT satellite systems. These indices provide near-real-time information about vegetation health and soil conditions. Combining satellite observations with meteorological data significantly improves forecasting accuracy.

The third innovation is its subregional approach. Rather than producing a single forecast for Northern Kazakhstan, the system generates forecasts at the district level, taking into account local climatic variability. Such geographically specific predictions are far more valuable for farmers and local authorities than broad regional forecasts.

In addition, the project compares the performance of five drought indices—SPI, HTC, SPEI, MCZI, and another widely used indicator—to determine which is most suitable for the climatic conditions of Northern Kazakhstan.

The project's main objective is to develop an early warning system for atmospheric droughts in the agricultural regions of Northern Kazakhstan by integrating climatic observations, satellite monitoring, and machine learning technologies. The system is designed to provide farmers and government authorities with timely and accurate information about impending drought conditions.

The research programme consists of three interconnected tasks.

The first task is to analyse long-term patterns of temperature, precipitation, and humidity and assess atmospheric drought risks in Northern Kazakhstan using meteorological data collected over the past 30 years.

The second task focuses on developing machine learning models for drought prediction by integrating meteorological observations with satellite-derived information on vegetation and soil conditions. GRU, Random Forest, and XGBoost algorithms will be trained and evaluated for this purpose.

The third task is to develop and test a fully operational early warning system in two regions of Northern Kazakhstan and prepare practical recommendations to help farmers and local authorities adapt to drought conditions.

Historical climate data—including air temperature, precipitation, and humidity—are obtained from the ERA5 reanalysis database of the European Centre for Medium-Range Weather Forecasts (ECMWF). Statistical analyses are performed in the R programming environment (RStudio) using the Mann-Kendall trend test, Sen's slope estimator, Run Theory, drought severity analysis, and correlation analysis.

Satellite observations from the Copernicus and EUMETSAT programmes are used to monitor vegetation and soil conditions through NDVI and EVI indices. Machine learning models are developed in Python using the NumPy library for data processing. GRU networks are applied to time-series forecasting, while Random Forest and XGBoost are used for drought classification. Model performance is verified through independent validation procedures.

During 2025, researchers collect and process thirty years of meteorological observations and conduct a comprehensive analysis of regional climatic conditions. The effectiveness of five drought indices is compared, while satellite data collection and vegetation monitoring begin.

In 2026, the machine learning models are trained, validated, and refined. Relationships between climatic variables and drought occurrence are identified, and forecasting accuracy is evaluated for both short-term and long-term predictions.

In 2027, the completed early warning system is implemented and tested in two regions of Northern Kazakhstan. Practical recommendations are prepared for farmers and local authorities, and the research findings are published in international peer-reviewed journals.

The project's final outcome will be a fully operational early warning system for atmospheric droughts, supported by an online information platform providing accessible data for farmers and government agencies. The system will enable advance planning of irrigation, adjustment of agricultural practices, and timely implementation of measures to reduce crop losses.

From a scientific perspective, the project will produce validated machine learning models with documented subregional forecasting accuracy and identify the most appropriate drought index for Northern Kazakhstan. The methodologies developed can subsequently be adapted to other regions with similar climatic conditions, both within Kazakhstan and internationally, giving the project broad scientific significance.

All research results will be published in journals indexed in Scopus and Web of Science, and presented at international scientific conferences.