Water demand is increasing worldwide, mainly owing to the growing population and industrial development [1]. Excessive population growth [2], accompanied by rapid industrialization and urbanization [3], has produced exceptionally high levels of water demand in developing countries. Climate change is another crucial issue that has affected water resources throughout the entire world. This is especially true for Mediterranean countries such as Turkey, as one of the world’s climate hotspots that will most likely become progressively drier and drastically warmer at higher levels of global warming [4]. Recent catastrophic heat waves, e.g., the blistering summer 2021 wave that resulted in wildfires, raged for nearly two months along the Akdeniz coast [5], decreasing river flows and reservoir levels [6], the emerging tragedy of degraded and drying lakes/wetlands, such as Lake Akgöl in eastern Turkey [7], Tuz Lake in central Anatolia plateau [8], and the Amik in southern Turkey [9], together with successive dry years, so that most of the country receives below average rainfall, all are the signals indicating that Turkey experiences intense drought [10]. Accordingly, many studies have investigated the spatial and temporal variation of drought hazards in Turkey in recent decades ^[10][11][10,11]. Overall, the studies confirm the fact that Turkey has been exposed to drought hazards rather frequently (every four or five years since the late 1980s). Intensive drought periods in 1971–1974, 1983–1984, 1989–1990, 1996–2001, 2007–2008, 2013–2014, and 2020–2021 have been reported ^[10][11][12][10,11,12], and projections have revealed increasing drought severity and frequency under the postulate climate change scenarios across the country ^[13][14][15][13,14,15]. In addition, an increasing number of studies have also attempted to develop statistical or dynamic models to anticipate droughts’ onset and severity, mostly in catchment scale. While some of these studies emphasize the interconnections between large-scale climate variables and severe drought events [16], the others highlight the capabilities of machine learning (ML) techniques to fit accurate nonlinear curves on the historical drought time series.

Drought is the consequence of a significant decline in the hydrological variables such as precipitation, soil moisture, and streamflow that undesirably affects all living beings. It is typically divided into four groups: meteorological drought (hereafter MD), agricultural drought (hereafter AD), hydrological drought (hereafter HD)MD, AD, HD, and SE ^[17][18][19,20]. All types start with MD, which is the result of precipitation deficiency. Dry spells in which rainfall amount is below the long-term precipitation average in a region lead to MD. MD causes the least disaster compared to the other types. For example, in case of insufficient rainfall, natural water sources such as lakes, wetlands, and groundwater reservoirs can be employed for water supply. HD is created by a decrease in precipitation and surface flow that might yield a decline in groundwater level. Since HD is closely related to water demand, it is crucial for urban catchments, industrialized regions, and agricultural activities. The AD, which is related to soil water storage and existing moisture capacity, is essential for crop growth and food security. In case of insufficient soil moisture for average crop growth and yields, AD may lead to crop failure, reduced range of productivity, livestock, and famine. The SD arising from the consequences of other types of droughts has important effects on society, and the economy can be expressed as a situation in which the demand for water exceeds the supply ^[17][19].

There are various indices for drought monitoring and assessment ^[18][20] that can identify the characteristics of drought, such as magnitude, severity, and duration. They are obtained from hydro-meteorological indicators, such as precipitation, temperature, runoff, soil moisture, reservoir storage, and groundwater level. Some indices are more appropriate than others for certain circumstances, such as the location of the study area, drought type, and availability of data. With the development of meteorological satellites and remote sensing (RS)RS technology on the one hand and the emergence of data-mining techniques on the other hand, a lot of current research has been conducted in the field of drought monitoring and forecasting (DMF) using these technologies. However, the relevant literature proved that the use of these tools is still not well known in some regions, and most of the studies related to DMF in those countries are based on observational ground data.

2.2. Meteorological Drought Indices

The Standardized Precipitation Index (SPI) suggested by McKee et al. ^[19][21] is a widely used MD index to assess and forecast precipitation deficit and surplus because of simplicity. The frequency of dry and wet spells at a specified time scale for any location can be determined by utilizing only long-term precipitation data (usually 30 years or more ^[20][21][22,23]. It is mainly advantageous that SPI can be calculated for diverse accumulation timescales (1, 3, 6, 9, 12, 24, and 36 months), so drought effect on the water resources depends on these timescales. For example, SPI at short timescales (1 to 3 months) may be associated with soil moisture, as SPI at relatively long timescales (>6 months) may be connected with groundwater, streamflow, and storage in reservoirs. SPI at a long timescale over 6 months can be a sign of hydrologic drought ^[22][23][24][24,25,26]. SPI calculation begins with the determination of the probability density function for long-term precipitation data by considering distribution fitting for each time scale. Although gamma distribution is generally used for SPI calculation, different distributions are used in the literature ^[25][26][27,28]. The most appropriate distribution can be selected considering the L-coefficients, skewness, and kurtosis ^[27][29]. The distribution function is calculated and normalized to obtain the standard normal random variable, in other words, the SPI value ^[17][19]. Although SPI values can be evaluated considering the different classifications, dry and wet periods are symmetric for all the classifications because of normalized SPI values ^[21][23]. Drought classes based on SPI include extremely wet (2.0 ≤ SPI), severely wet (2.0 < SPI < 1.5), moderately wet (1.49 < SPI < 1.0), near normal (−1.0 < SPI < 1.0), moderate drought (−1.49 < SPI < −1.0), severe drought (−2 < SPI < −1.5), extreme drought (SPI ≤ −2.0) ^[19][21]. The Standardized Precipitation Evapotranspiration Index (SPEI) is a rather newly developed MD index ^[28][30] based on both precipitation and evapotranspiration. Although SPEI calculation is as simple as SPI calculation, it is a multi-scalar index including climatic water balance, which is the difference between precipitation and evapotranspiration. Thus, it is more sensitive to changes in temperature ^[29][31]. Although evapotranspiration is usually calculated by the Thornthwaite method because of its simplicity and requirement for few data such as latitude and daily temperature, as was suggested in the original SPEI calculation, some research indicated that this method underestimated or overestimated the evapotranspiration. So, the Penman–Monteith or the Hargreaves equations can be used to obtain evapotranspiration. While daily maximum and minimum temperature data is needed for the Hargreaves equation, the Penman–Monteith requires relatively more long-term data such as temperature, wind speed, relative humidity, and solar radiation. A comparison of drought classes between SPI and SPEI was presented by Danandeh Mehr et al. [15]. The De Martonne Aridity Index (DAI) is calculated utilizing monthly temperature and precipitation data. Annual values are also used to classify regional climate and temporal changes while defining aridity as the ratio of precipitation (mm) to mean temperature (°C) ^[30][32]. The DAI is always positive, and the seven drought classes according to DAI are defined: arid (DAI < 10), semi-arid (10 ≤ DAI < 20), Mediterranean (20 ≤ DAI < 24), semi-humid (24 ≤ DAI < 28), humid (28 ≤ DAI < 35), very humid (35 ≤ DAI ≤ 55), and extremely humid (DAI > 55) ^[31][33]. The De Martonne–Gottman index (DMGI) is a modified DAI index that is obtained using the precipitation and temperature at the driest month as well as mean annual temperature and total annual precipitation. Climate characteristics based on DMGI are indicated: polar, desert, steppe semi-arid, between steppe and moist, steppe semi-humid, humid, very humid, wet for the DAI values of below 0, 0–5, 5–10, 10–20, 20–28, 28–35, 35–55, and higher than 55, respectively ^[32][34]. The Palmer drought index (PDSI), which was developed by Palmer ^[33][35] and Alley ^[34][36], is based on a soil–water balance equation. Thus rainfall, runoff, soil moisture, and potential evaporation are required. PDSI is very closely affected by calibration periods. The availability of calibration is limited outside the calibration area. This index is not spatially comparable. Due to the disadvantages of PDSI mentioned in the previous sentence, the Self-calibrated Palmer Drought Severity Index (sc-PDSI), which is spatially comparable, has been developed, but even so, the major disadvantages associated with fixed temporal scale and exposure from the cases up to the preceding four years have not been overcome. PDSI values may usually change between −6 and +6, and the drought classification is as follows: near normal (−0.5 < PDSI < 0), incipient drought (−1.0 < PDSI < −0.5), mild drought (−2.0 < PDSI < −1.0), moderate drought (−3.0 < PDSI < −2.0), severe drought (−4.0 < PDSI < −3.0), extreme drought (PDSI ≤ −4.0) ^[35][37]. If soil moisture and/or the lakes, rivers, and reservoirs are significantly different from normal levels after drought termination or the end of the wet spell, it can be more appropriate to use Palmer Hydrological Drought Index (PHDI) depending on precipitation and soil moisture storage instead of meteorological index, PDSI ^[36][38]. Since only the rain is accepted as precipitation, not snowfall, snow cover, and frozen ground for PDSI and PDHI, this situation may lead to the wrong timing of the indices. The China Z index (CZI) was introduced by the National Climate Centre of China for drought monitoring. CZI only requires monthly precipitation data and is recommended over SPI because it is easier to calculate in case of a lack of precipitation data ^[37][39]. The Pearson Type III distribution is the best-suited distribution for precipitation. Hence CZI depends on the Wilson–Hilferty cube-root transformation from the chi-square variable to the Z-scale ^[38][40]. The Modified Chinese Z index (MCZI) uses the data’s median for precipitation rather than the mean difference from the CZI calculation ^[39][41]. The Rainfall Anomaly Index (RAI) is a calculated comparison between each precipitation data and the ten lowest or ten highest precipitation data considering the long-term mean of precipitation ^[40][42]. The RAI can be used to determine various accumulation times, such as 1, 3, 6, 9, and 12 months and be used under circumstances in which 30 years of precipitation data are not available, and precipitation is not normally distributed ^[41][43]. The Deciles Index (DI) only requires precipitation data. Precipitation data should be normalized if it does not fit the normal distribution. Precipitation data sorted in ascending order is separated into ten groups, normal distribution deciles distribution, and cumulative frequency distribution is obtained. The first decile consists of the precipitation data less than 10% of the total amount of precipitation. The next decile involves the data from 10% up to 20%, and so on for the other deciles. DI is calculated by dividing the cumulative probabilities into five classes which consist of 20% (two deciles) in ascending sort: much below the normal, below the normal, near the normal, above the normal, and much above the normal, respectively. Drought severity is calculated by comparing the precipitation of a specified month with long-term cumulative distribution precipitation ^{[39][40][41][42]}[41,42,43,44]. The Reconnaissance drought index (RDI) is stated as the initial value (α), normalized RDI (RDI_n), and standardized RDI (RDI_st). The initial value is obtained as the ratio of total precipitation to total potential evapotranspiration, RDI_n is computed using the arithmetic average of the initial values, and RDI_st is generated by the same method as SPI, so thresholds are the same for both indices ^[43][45]. Since the formulation of RDI_st consists of the natural logarithm of the initial value, RDI_st cannot be calculated if the total precipitation is zero. For this situation, different procedures have been suggested ^[44][46]. Because RDI does not depend on the potential evapotranspiration (PET) calculation method, RDI can be reliably evaluated even if only precipitation and temperature data are available ^[45][47]. The Percent of Normal Index (PNI) is calculated by the percentage ratio of the actual precipitation average of the period ^[46][48]. This period can be a month, rainfall season, or 12 months (annually). It is necessary to have at least 30-year precipitation data ^[47][49]. Besides, there are drawbacks that precipitation does not fit the normal distribution, and PNI depends on the location and season, so it is effective for a single region or season ^[48][50]. The Z-Score Index (ZSI) is obtained by the ratio of subtraction of the average from the actual precipitation to the standard deviation. Higher values mean a higher severe drought ^[38][40]. Since it does not need to adjust any distribution, such as Pearson III or Gamma, for the ZSI calculation, it may not be representative of shorter timescales compared to SPI. In addition, missing values in precipitation data are not a problem in this method, such as CZI ^[37][39]. The Erinç drought index (EDI)is calculated by dividing total precipitation by the average maximum temperature. Then, it is classified according to the areal distribution of vegetation formations in Turkey, as follows: severe arid (<8), arid (8–15), semi-arid (15–23), sub-humid (23–40), humid (40–55), per humid (>55) ^[49][51]. In order not to predict a more humid climate than it is in continental, cold winter or very cold climatic regions such as the northeastern Anatolia (Erzurum-Kars) part of Turkey, Erinç has adapted to use the average maximum temperature values instead of the mean temperature in the calculation of the index. Erinç suggested that the index should not be used in months when the mean maximum temperature value falls below 0 °C when evapotranspiration cannot occur ^[50][52].

2.3. Agricultural Drought Indices

The Normalized Difference Vegetation Index (NDVI) was suggested by Rouse ^[51][53] in 1974 as a vegetation index based on data from RS. NDVI is used to quantify vegetation greenness and vegetation density and assess changes in plant health. Visible wavelengths between 390 nm and 700 nm, especially the red wavelengths between 620 nm and 700 nm, are absorbed by the pigments in plant leaves during photosynthesis, while the near-infrared wavelengths within the range from 760 nm to 900 nm are reflected by spongy mesophyll in the plant. Considering the normalization of this difference for different ranges using satellite images enable us to find the NDVI ^[52][53][54,55]. The pattern of the vegetation change can be determined by comparing the NDVI values obtained from satellite images at different times. High values of NDVI, ranging from −1 to 1, account for higher biomass, vegetation, and vegetation density. While low vegetation and bare soil correspond to a negative value or close to zero, water, clouds, and snow indicate low values of NDVI. The areas with a low NDVI value where agriculture is intensive indicate poor plant growth owing to several causes, for instance, redundant moisture, drought, pests, or disease ^[53][55]. The Vegetation Condition Index (VCI) is based on NDVI images considering vegetation cover. VCI is expressed by a percentage by dividing the subtraction of the minimum value from the value of VCI at a given pixel by the difference between the maximum and minimum of VCI and multiplying by 100. While VCI higher than 50% represents more plant growth, dry conditions, and low vegetation growth are associated with VCI values less than 50 and approaching zero ^[54][56].

2.4. Hydrological Drought Indices

The most commonly utilized HD indices are the Standard Streamflow Index (SSI), Streamflow Drought Index (SDI), and Standardized Runoff Index (SRI). The main difference between the SSI calculation and the SPI calculation is that flow is employed in place of precipitation. Drought classification based on the thresholds is the same as SPI ^[55][57]. Since the mean and standard deviation of the standardized variable SSI are 0 and 1, respectively, SSI can be evaluated temporally and spatially ^[56][58]. Using the same methodology as SPI calculation, the SDI ^[57][59] is computed for the hydrological year starting in October and ending in September. However, cumulative streamflow volume rather than rainfall data is utilized for SDI. Negative SDI values are evaluated as dry periods, the classification based on SDI as follows: mild drought (−1.00 ≤ SDI < 0.0), moderate drought (−1.5 ≤ SDI < −1.0), severe drought (−2.00 ≤ SDI < −1.5), and extreme drought (SDI < −2.0). The SRI ^[58][60] used runoff data for the calculation, the same as the SPI, and can be investigated by the HD temporally and spatially. The drought classification of the SRI values is the same as the classification of SPI ^[59][61].

2.5. Socioeconomic and Ecological Drought Indices

The Multivariate Standardized Reliability and Resilience Index of Socioeconomic Drought (MSRRI) proposed by Mehran et al. ^[60][62] is combined the inflow-demand reliability (IDR) index and the water storage resilience (WSR) index by a statistical approach. While IDR is based on the top-down methodology evaluated with available water resources corresponding to the water demand, WSR is related to bottom-up as sufficient reservoir storage for the upcoming water demand. The socioeconomic drought index (SEDI) is determined according to water shortage level at reservoirs and drought duration ^[61][63]. It has four levels: Level 1 for slight, 2 for moderate, 3 for severe, and 4 for extreme. Since SEDI considers the inconsistent situation with the reality that all prior water deficits must be made up with surplus water later, SEDI overestimated the impact of the socioeconomic drought. Therefore, the water resources system resilience (WRSR) index was suggested that considers the percentage of recoverable part of antecedent water deficit from excess water in later periods ^[62][64]. As the WRSR value, which ranges from 0 to 1, increases, a more severe socioeconomic drought can be expected. Standardized supply and demand water index (SSDWI) is another socioeconomic drought index ^[63][65] that is calculated by the difference function using monthly water supplies and demands within a basin, run theory, and copula functions for analyzing drought characteristics at different timescales.

2.6. Drought Monitoring

Dabanlı et al. ^[64][18] examined the spatial and temporal patterns of drought to monitor the variability of drought across Turkey. To this end, precipitation records were obtained from 250 meteorological stations for the period between 1931 and 2010 (80 years). The drought analysis was performed using SPI values at 1-, 3-, 6-, 9-, and 12-month timescales. The SPI time was divided into two groups based on principal component analysis PCA, and then Turkey was classified into four regions based on the classified SPI groups. The temporal variability of drought events and their cyclical patterns were examined in each region. The results showed that widespread drought events were experienced in 1970, 1990, 1973, and 1957 for four regions of Turkey based on SPI-3, whereas drought events were observed in 1990, 2000, 1957, and 1955 for four regions of Turkey based on SPI-12. Moreover, the periodic characteristic of drought patterns showed that the frequency of short-term drought ranges between 2 and 5 years across all regions. However, the frequency of long-lasting drought varies between 10 and 20 years. Dursun and Babalık ^[32][34] investigated the historical drought events in Isparta, Antalya. The precipitation and temperature observations were gathered from Atabey, Eğirdir, Isparta, Senirkent, Uluborlu, and Yalvaç meteorological stations between 1990 and 2020. DMGI and SPI series were calculated to monitor drought events in the region. The findings demonstrated that near-normal droughts were identified at all the chosen locations. The longest drought duration was detected between September 2004-March 2010 (67 months) in Yalvaç station, considering SPI-12 values. Dikici ^[53][55] investigated the drought events using vegetation indices such as NDVI and VCI to monitor drought in the Seyhan basin. NDVI and VCI values were attained using RS data. In addition, three MD indices, including DI, SPI, and SPEI, and one HD index (i.e., SRI) were adopted for comparison with AD. The results of the vegetation indices showed that the Seyhan basin experienced drought events in 1973–1974, 1989, 2001, 2007–2008, 2014, and 2016. An increasing trend in drought events was detected in the region. It was observed that other indices, such as DI, SPI, SPEI, and SRI, had been determined to be more accurate and to have a significant correlation with each other based on historical data. Yıldız ^[65][67] investigated spatiotemporal drought variation over the CAR for the period of 1953–2004. To this end, historical SPI time series from 27 meteorology stations were used. A drought intensity–areal extent frequent curve for the area was created using monthly SPI values. Severe drought events were seen in the area during the years 1956, 1961, 1964, 1973, 1977, 1984, 1989, 1993–1994, 199–2001, and 2004. It was found that historical droughts of 1956, 1964, 1984, 1993, 2001, and 2004 have a range of return periods ranging from 2 to 25 years. More than 40% of severe droughts were observed between 1956 and 2001 in the CAR. Moderate drought events occurred during all drought years for more than half of the region.

2.7. Drought Forecasting

Drought forecasting is a crucial part of managing and planning water resources. Although it is extremely difficult to predict when the drought will begin and end, numerous droughts forecasting models have been created to increase the ability to predict droughts. ^[66][67][107,108]. The LR analysis and artificial intelligence-based models (e.g., Artificial Neural Networks (ANN), Support Vector Machine (SVM), Adaptive Neuro-Fuzzy Interface System (ANFIS), Decision Tree (DT), Random Forest (RF), Genetic Programming (GP)) are of the most widely used ML methods for drought forecasting. Hybrid models, which are the combinations of more than one ML technique, are another useful drought forecasting methodology. Several drought forecasting models across Turkey have been published by researchers in the relevant literature. For instance, Başakın et al. ^[68][93] developed SVM and K-nearest neighbor (KNN) models to forecast one, three, and six-month ahead PDSI values which were acquired from https://crudata.uea.ac.uk/cru/data/drought/ (accessed on 25 September 2018). In addition, the wavelet transform (WT) technique was integrated with each modeling approach to improve forecasting accuracy. To this end, 116 years of PDSI values were collected for Kayseri Province. RMSE and Nash–Sutcliffe Efficiency (NSE) were used to evaluate the generated model’s performance. The results showed that SVM and KNN models could be used for drought forecasting. However, it was observed that wavelet transform integrated with these models could improve the forecasting performance substantially. Tufaner and Özbeyaz ^[69][94] collected monthly averages of temperature, pressure, relative humidity, wind speed, runoff, and total precipitation data for Adıyaman Province. The study made use of the Adıyaman Province’s meteorological data covering the period from 1980 to 2011. Then, PDSI values were calculated and used for drought forecasting. LR, DT, SVM, and ANN models were developed to predict PDSI values. Several performance evaluation criteria were adopted to compare the forecast performance of developed models. For PDSI forecasting, the study showed that the ANN model performs better than the others. Mehdizadeh et al. ^[70][98] developed several drought modeling approaches using classic time series models such as linear autoregressive and nonlinear bi-linear in Turkey. Moreover, hybrid models were created by a combination of WT and gene expression programming (GEP) using five different mother wavelets. Three different timescales of SPEI, such as SPEI-3, SPEI-6, and SPEI-12, were calculated using the data gathered from six meteorology stations in Ankara. The created model’s performance was evaluated in terms of mean absolute error, RMSE, and R indicators. The findings of the model comparison revealed that bi-linear models are more accurate than autoregressive.