Nineveh Province is located between latitudes 34° 52′ 25″ – 37° 06′ 3.5″ N and longitudes 41° 10′ 21″ – 43° 42′ 50″ E and has an area of 35,152 km2. It is bordered to the north by Duhok Governorate, east by Erbil and Kirkuk Governorates, south by Salah al-Din and Al-Anbar Governorates, and west by the Syrian Arab Republic, as shown in Figure 1 (Mohammed et al., 2022). Iraq is located within the arid and semiarid climatic zone and is characterized by the succession of four climatic seasons: summer (July–October), hot and dry; autumn (October–November); winter (December–February), temperate and cold; and finally, rainy (rain). Approximately 90% of the total annual rainfall is concentrated during the winter and spring (Salman et al., 2019). As shown in Table 1, the locations of the selected stations in the study area and their altitudes above sea level vary from station to station due to the topography of the study area.

Source: Republic of Iraq, Ministry of Transport. Iraqi Meteorological Organization/Climate Department, unpublished data.

Topographic features are among the most important natural factors affecting the climate of the study area. The diversity of topographic features, such as mountains, plateaus, hills, and valleys, and their lack of uniformity, create climatic variations from one region to another (Farouk et al., 2005). Nineveh Governorate is characterized by diverse surface features. The topography and elevation above sea level are what control the elements of the climate. The most important topographical feature of the study area is the existence of two regions: the mountainous region and the undulating region and plains.

This study relies on the application of an integrated set of advanced technical and statistical tools to analyse climatic data and to forecast rainfall amounts with a high degree of accuracy, as these tools encompass a range of analytical and predictive modeling techniques.

The data represent the recorded rainfall amounts over the period from 1994 to 2024 at a set of meteorological stations distributed within the study area. These data were adopted as the basis for the analysis of rainfall characteristics and the construction of predictive models. This period was adopted due to the limited availability and discontinuity of data in earlier years, which could affect the accuracy of the analysis. Accordingly, the relatively complete available records were used as a reliable basis for analyzing rainfall characteristics and developing predictive models. Rainfall is one of the most important forms of condensation in the upper layers of the atmosphere. It depends on the amount of water vapor present and the temperature of the dew point at which condensation occurs. Rainfall is one of the most important forms of atmospheric condensation. Its occurrence depends on the amount of water vapour in the atmosphere and the attainment of the dew point temperature.

Iraq is located within the arid and semi-arid climatic zone, and its rainfall regime is influenced by the Mediterranean rainfall system. Consequently, rainfall is concentrated during the cold season, while the hot season is generally dry (Al-Mousawi et al., 2013). Table 2 shows that rainfall in Nineveh Governorate has a clear seasonal character, with precipitation concentrated from January to April, with a pronounced peak in January, February, and March across all stations (Al-Azzawi, 2017). Rainfall values reach monthly averages of approximately 60-70 mm at some stations, particularly at the Makhmour, Rabia, and Sinjar stations. Rainfall continues during April in moderate amounts and then begins to decrease sharply in May. The region enters a period of near-total drought during the summer (June-August), as the values are almost nonexistent at all stations. With the beginning of autumn (September-October), the rains gradually returned. For the annual total, the highest values were recorded at the Rabia (365.5 mm) and Makhmour (364.6 mm) stations, whereas the lowest values were recorded at Tel Abta (253.2 mm), reflecting spatial variation associated with geographical location and topography.

Source: Based on the Ministry of Transport, Iraqi Meteorological Organization and Seismological Monitoring, Climate Department, (unpublished data), 2024.

The construction of the predictive model involved incorporating a set of climatic variables that affect rainfall amounts, including as shown in Table 3 temperature, relative humidity, wind speed, and atmospheric pressure, as key explanatory variables in estimating rainfall behavior.

Source: Based on the Ministry of Transport, Iraqi Meteorological Organization and Seismological Monitoring, Climate Department, (unpublished data), 2024.

The program was based on Python as the main programming language to utilize the artificial intelligence libraries and data analysis tools (Aurélien Géron, 2019). The pandas library was used for processing and organizing climate data (Reback et al., 2020). Calculations and statistical analysis were also done using the NumPy library (Harris et al., 2020). The predictive model performance was evaluated using the Scikit-learn library (Pedregosa et al., 2011). In addition, the Matplotlib and Seaborn libraries were used to create the graphs, analyse climate patterns, and present the findings in an understandable graphic format (Hunter, 2007).

The first stage involved processing climate data by assessing data quality and filtering out extreme and illogical values, particularly anomalous or negative rainfall values, and addressing missing values to ensure the consistency and accuracy of the time series data (Han et al., 2012). Owing to the prevailing climate in the study area, the summer months (June, July, and August) were excluded from the analysis because they represent a period of near-total drought. The monthly totals for the climatic variables, including temperature, relative humidity, wind speed, and atmospheric pressure, were then calculated. The relationships between independent climate variables and rainfall amounts were examined to identify the most influential variables and reveal patterns of correlation among them, thereby improving the efficiency of the model and enhancing the accuracy of the predictive results (Wilks, 2011).

The first stage involves processing climate data by assessing its quality and filtering out extreme and illogical values, especially anomalous or negative rainfall values, and addressing missing values to ensure the consistency and accuracy of time series data (Han et al., 2012). Owing to the prevailing climate in the study area, the summer months (June, July, and August) were excluded from the analysis because they represent a period of near-total drought. The exclusion of the summer months (June-August) was primarily intended to avoid introducing noise associated with near-zero rainfall values, which could adversely affect model training and reduce predictive stability, particularly in data-driven approaches. The monthly totals for the climatic variables, including temperature, relative humidity, wind speed, and atmospheric pressure, were then calculated. The relationships between independent climate variables and rainfall amounts were examined to identify the most influential variables and reveal the patterns of correlation between them, which contributed to improving the efficiency of the model used and enhancing the accuracy of the predictive results (Wilks, 2011).

To investigate the relationships between rainfall and the associated climatic variables, correlation analysis was performed using both Pearson’s and Spearman’s correlation coefficients. Pearson’s correlation coefficient was applied to evaluate the strength and direction of linear relationships among the climatic variables) Benesty et al., 2009), as expressed in Equation (1). This method was used to identify variables that exhibit strong linear associations with rainfall patterns.

where (r) is Pearson’s correlation coefficient, (Xi) and (Yi) represent the observed values of the climatic variables, and (n) is the total number of observations.

In addition, Spearman’s rank correlation coefficient was employed to assess monotonic relationships that may not necessarily follow a linear pattern (Spearman, 1904), as shown in Equation 2. This approach improves the robustness of the statistical analysis by detecting non-linear associations among climatic variables before their integration into predictive machine learning models.

Where ( ) represents Spearman’s rank correlation coefficient, (di) is the difference between the ranks of paired observations, and (n) is the number of observations.

This study relied on a set of climatic variables to train and evaluate the models: temperature, relative humidity, wind speed, and atmospheric pressure. Given the correlation between rainfall amounts in the study area and these variables, rainfall data recorded at seven spatially distributed meteorological stations within the study area were utilized (Gu et al., 2022). Approximately 90% of the data from each station were allocated for model training, while the remaining 10% were used to evaluate its predictive efficiency. The study was based on monthly averages of climatic elements. The climatic data spanning 31 years for each station were divided into two independent groups. The first group included training data for the period (1994-2014), whereas the second group included test data for the period (2015-2024) , The period (2015-2024) was used as an independent test phase to evaluate the ability of the trained models to predict observed rainfall before the validated model was applied to future forecasts during the period (2025-2034) (Hastie et al., 2009).

The performance of the predictive models was evaluated on the basis of test data not included in the training process, ensuring the neutrality of the evaluation and the accuracy of measuring the predictive ability of the models. Through the trained models, future forecasts of rainfall amounts for the period from 2025-2034 were obtained (Kuhn et al., 2013). The study methodology is illustrated in Figure 2.

To detect monotonic long-term trends in rainfall time series (1994–2024), the non-parametric Mann–Kendall (MK) test was applied, as it is widely used in hydro-climatological trend detection studies for its robustness to non-normal data and outliers (Hamed, 2008).

The magnitude of change associated with the detected trend was quantified using Sen’s slope estimator, which provides a median-based estimate of the annual rate of change (mm/year) without assuming a specific data distribution (Sen, 1968).

The combined use of the Mann–Kendall test and Sen’s slope estimator has become a standard approach in contemporary climate trend assessments, particularly in semi-arid environments characterized by high rainfall variability (Yue and Wang, 2004; Hamed, 2008; Tabari et al., 2018). A significance level of α = 0.05 was adopted to determine the statistical significance of detected trends.

This study was premised on the creation of an extremely precise predictive model to approximate the rainfall quantities in the research site by implementing an integrated approach that incorporates statistical modeling and contemporary machine learning (Khan et al., 2023). The models have had some success in explaining the temporal relationships among various climatic variables and rainfall quantities, which has given them a strong methodological basis to proceed to more sophisticated predictive approaches and improve the fidelity of rainfall forecasting (Ghosh et al., 2023). The methodological framework of the study was extended with the use of advanced machine learning models that were selected as integrative models and incorporated the results of another set of predictive models . These included:

1. XGBOOST

The XGBoost algorithm is an advanced gradient boosting technique based on the Gradient Boosting Machine (GBM) framework. It was introduced by (Chen and Guestrin, 2016) and has demonstrated high performance in supervised learning tasks, including regression, classification, and ranking problems. The algorithm constructs an ensemble of decision trees iteratively, where each new tree minimises the prediction error of the previous trees through gradient optimisation. XGBoost also incorporates regularisation terms to improve model generalisation and reduce overfittingm, making it highly effective for rainfall prediction modeling.

In this study, the XGBoost model was applied to predict rainfall patterns by minimizing the objective function, which consists of a loss function and a regularisation component, as expressed in Equation 3:

where represents the objective function, denotes the loss function measuring the difference between observed and predicted rainfall values, represents the individual decision trees, is the regularisation term used to control model complexity, and (K) is the total number of trees in the ensemble.

2. CatBoost

The CatBoost algorithm is a modern incremental boosting algorithm that was created by Yandex (Saber et al., 2022). A key feature of this algorithm is that it can directly operate with categorical data, which is a frequent issue in most machine learning problems. Conventional approaches may need elaborate and time consuming preprocessing steps that may implicate data leakage. The CatBoost algorithm on the other hand directly works with categorical variables in its own architecture simplifying the process of data preparation to a large extent. The algorithm is efficient and simple to use since it is highly accurate and resistant even with default parameter settings. CatBoost is very efficient in most machine learning tasks, such as regression, classification, and ranking, thus it is an apt solution to working with complicated nonlinear data in predictions.

3. Random Forest

The Random Forest algorithm, introduced by Breiman (2001), is an effective ensemble machine learning algorithms for regression modeling and classification. This study used the random forest algorithm to predict rainfall amounts. The algorithm relies on creating many decision trees and standardizing their outputs. The final prediction is calculated on the basis of the average of the tree predictions in the case of regression (Sharma et al., 2021), which contributes to improving accuracy and reducing the problem of overfitting. Random Forest is particularly effective in modeling nonlinear relationships between climatic variables and rainfall patterns. It also demonstrates strong performance when handling large datasets and missing values, making it suitable for climatic forecasting applications.

The Gini impurity criterion used in decision tree splitting is expressed in Equation 4.

where (Pi) represents the probability of class (i), and (c) denotes the total number of classes. Lower Gini values indicate higher node purity during the tree construction process.

4. LSTM Neural Network

Neural networks Advanced types of recurrent neural networks include long Short-Term memory(LSTM) networks. They were devised by Hochreiter and Schmidhuber 1997) to overcome the vanishing gradient issue that conventional models have when applied to long time series. This study paper integrated an LSTM network to predict precipitation levels in time series and model time series data by taking advantage of the capacity to represent time patterns of complex time variations and nonlinear interactions among climatic factors. Past research has shown that LSTM models are very efficient in hydrological and climatological applications, especially in predicting precipitation and surface runoff, as compared to conventional statistical models (Kratzert et al., 2018).

Wolpert (1992) introduced the stacking strategy for ensemble learning. It is one of the more complex approaches to machine learning that leverages the interdependence among a collection of underlying models to enhance predictive capacity and generalizability. In this study, four base models (XGBoost, CatBoost, random forest, and an LSTM neural network) were used to leverage each model's strengths in capturing nonlinear relationships and temporal dependencies in climate data and, subsequently, enhance the precision of rainfall predictions. The two stages in this model mechanism are dependent. The initial step involves training the four base models separately on climatic element data and extracting each model's predicted value. The second step entails integrating the outputs using a stacked ensemble model to produce the final rainfall forecast. To prevent the issue of overfitting, the meta-model is not trained with the initial outputs of the base models but instead is trained based on a cross-validation strategy through a leave-one-out approach. The results from the cross-validation folds are aggregated to create a new dataset, which is used to train the meta-model; this is why this strategy is known as stacking. Combining basic models is an essential step in creating a stacking model, as various machine learning models can serve as meta-models, including linear regression and random forests. The weighted combination method was adopted in this research. The contribution of each basic model was assigned different weights in the final rainfall prediction and can be mathematically represented by equation 5.

where ωm (m = 1, 2, …, M) is the weight assigned to each base model and fm,i represents the prediction of model m for the ith observation.

To obtain the optimal final prediction, the set of stacking weights was estimated by minimizing the mean squared linear regression. Thus, the objective function under two constraints is as in Equation 6.

where (Yo,i) is the observed rainfall value for observation (i), and (N) represents the total number of observations.

The optimization process was performed under the following Equations 7 and 8.

where Ω= [W1,W2,…, Mw]. The sum of the weights assigned to the base models is subject to two constraints: the weights must be nonnegative, and their sum must equal one. This leads to a quadratic optimization problem, which was solved via the qpsolvers package in Python. By calculating the weights of the base models, the final rainfall prediction was integrated into the stacking ensemble model.

A comparison of the accuracies of the models before and after parameter optimization in Table 4 reveals a clear improvement in the performance of the models used in rainfall prediction. The stacking ensemble (Stacking Ensemble) model achieves the highest accuracy after optimization, reaching 98.10% compared with 95.9% before optimization, which confirms the effectiveness of integrating the basic models in enhancing the predictive capability. The XGBoost and CatBoost models also significantly improved in accuracy after optimization, whereas the accuracy of the LSTM model markedly increased because of its efficiency in representing the time dependence of climate series. In contrast, the random forest model showed a slight decrease in accuracy after optimization. This indicates its sensitivity to parameter adjustment and, in general, confirms that the stacked ensemble model outperforms individual models in achieving more accurate rainfall predictions in the study area.

The performance evaluation results in Table S1 show the superiority of the stacking model over the individual models at all the study stations, as it has the lowest error metric values ​​and the highest coefficient of determination values ​​compared with the CatBoost, Random Forest, XGBoost, and LSTM models when applied individually. This superiority reflects the effectiveness of integrative integration in leveraging the different strengths of each algorithm and improving the predictive capability of the final model (Latif et al.,2023).

In terms of the performance evaluation results, the stacking model achieved the best overall performance compared with the individual models at most stations. The Al-Baaj station recorded the lowest error values (RMSE=0.7 and MAE=0.5) with the highest explanatory power (R²=0.8). similarly, at Tal Abtah , Rabia, and Makhmur , the stacking model showed notable improvement over the other models, achieving relatively high R² values reaching up (0.7,0.9,and 0.9) respectively , with comparatively lower metrics. Advanced results were also achieved at the Tal Abtah, Rabia, and Makhmur stations, with R² values of approximately 0.8 and a relative decrease in error measurements. In contrast, the Tal Afar station presented the highest error values and the lowest relative stability of the models, as the RMSE values for the stacking model reached approximately (1.2) and (MAE = 0.8), reflecting the complexity of rainfall behavior in this region. The average performance of the Sinjar and Mosul stations was also recorded, with RMSE values ranging from 0.9 and R² values ranging from 0.7.

Table 5 displays the Pearson correlation coefficient of Pearson, which indicates depicts that the linear relationships existing between the climatic elements within the study area are very strong. The amount of rainfall had a strong negative relationship with maximum and average temperatures (r =-0.7) and a moderate negative relationship with minimum temperatures (r =-0.5) which means that higher temperatures are linked to a significant reduction in precipitation. Rainfall on the other hand had a strongly positive association with relative humidity (r = 0.7) and atmospheric pressure (r = 0.6) which is characteristic of the physical nature of the precipitation regime in semi arid conditions where cool and moist air masses are correlated with more rainfall. The outcomes also depicted a negative association that is apparent between rainfall and hours of sunshine (r = -0.6) and the converse of the cloud cover and the sunshine.

Wind speed, however, was strongly correlated with the rainfall (r = 0.1), which means that it is a rather weak factor that affects the precipitation levels directly. Very high correlation coefficients were obtained at the level of interaction of independent variables between Tmax and Tmean (r = 0.99) and between Tmin and Tmean (r = 0.99), between temperatures and hours of sunshine (r = 0.99), and at the same time, a strong negative correlation was observed between heat and humidity (r = -0.9). Such large values suggest that there is a high level of multicollinearity among the variables of the thermal variables, and these variables should be considered when constructing predictive models.

Spearman's results in Table 6 confirmed the existence of strong, consistent, and even stronger relationships with Pearson's findings, indicating that the relationships are not only linear but also monotonic. Rainfall showed a strong negative correlation with all temperatures (ρ = −0.8). A very strong positive correlation was observed with relative humidity (ρ = 0.9) and atmospheric pressure (ρ = 0.8), reinforcing the hypothesis that humidity and pressure are crucial factors in shaping precipitation behavior. A strong negative correlation was also found between rainfall and hours of sunshine (ρ = −0.8), confirming that periods of high radiation coincide with low precipitation. Wind speed, however, had a relatively weak effect (ρ = −0.2). Spearman's coefficients show nearly perfect values ​​(ρ = 1.0) between Tmax, Tmin and Tmean, indicating near-perfect monotonic homogeneity between the temperature variables, and confirming the existence of a strong intrinsic dependence between them.

In general, Pearson and Spearman's results reveal a strongly interconnected climate system governed by a clear thermal-humidity-pressure triad; precipitation decreases as temperatures rise and solar radiation increases. This increases with rising humidity and atmospheric pressure. The results also confirm a high degree of internal correlation among the thermal variables, necessitating caution when incorporating them into predictive models to avoid multicollinearity. The results also show very high correlations between maximum, minimum, and average temperatures (≈ 1.0), indicating strong linear overlap between them. Therefore, including these variables together in the predictive model may lead to multicollinearity, necessitating the selection of one variable or addressing the relationship between them to ensure model stability and accuracy.

Based on the results of correlation analysis (Pearson and Spearman), the climatic variables included in the forecasting models were systematically selected, taking into account the strength of the relationship with rainfall amounts and trends. The results showed very high correlations between temperature variables (maximum, minimum, and average). This suggests the possibility of multicollinearity among them. Therefore, these variables were carefully managed during model construction, either by selecting a representative variable or by addressing the interrelationships between them, to ensure model stability and improve its explanatory and predictive capabilities.

The outcomes of the Mann-Kendall test on the annual rainfall totals during the period (1994-2024) reveal that there is no statistically significant monotonic trend at a significance level of 0.05 at each of the studied stations, as presented in Table 7. The p-values were much larger than the critical value. This is indicative of the lack of a long-term, systematic change in the amount of precipitation per year in the governorate over the period of study. Even though the trend was not statistically significant, Sen’s slope estimator values by Sen varied in signal and magnitude among the stations. The Tel Abtah and Sinjar stations had relative annual growth rates of about 2.8 and 2.76 mm/year, respectively. Whereas the Mosul and Al-Baaj stations showed slight negative tendencies of approximately −1.34 and −1.33 mm/year, respectively. the other stations portrayed a minimal positive tendency of about 1 mm/year.

These values ​​indicate that the observed changes, although spatially different in direction, remain small in magnitude and insufficient to produce a structural shift in the annual rainfall pattern within the time frame studied. This reinforces the hypothesis of relatively stable annual rainfall in the semi-arid environment of Nineveh Governorate, with natural oscillations between years remaining the dominant feature of the time series.

Table S2 shows the results of the Mann-Kendall test and Sen’s slope estimator for seasonal rainfall trends at the study stations during the period (1994–2024), where the results generally show weak temporal trends and spatial instability, as the majority of values ​​did not register statistical significance at the (α = 0.05) level. Autumn was characterized by slight, mostly positive, but low values ​​that do not reflect actual climate change, while spring showed relatively higher rates of positive trend at some stations without reaching the significance level Statistically, while the summer season was characterized by an almost complete absence of trend with a slope close to zero at all locations, the winter season saw some stations record decreasing slopes of varying magnitudes, but these were not statistically significant. Overall, the quarterly results reflect a natural annual fluctuation rather than a long-term, regular climatic shift in the rainfall regime of Nineveh Governorate.

Table S3 of the monthly analysis shows that statistically significant trends were concentrated to a limited extent in the spring months, particularly April and May at some stations. While the summer months were characterized by an almost zero trend, the winter and autumn months showed mostly inconsistent and insignificant trends, confirming that the observed change reflects a normal seasonal fluctuation rather than a long-term, regular monthly climate shift.

Comparisons of the historical rainfall record with the forecast outcomes of the period (2025 2034) at the Nineveh Governorate stations revealed that the overall seasonal trend of the rainfall was always the same, unlike the fact that there were no clear qualitative changes in monthly means and annual totals as indicated in Tables 2 and 8, respectively. In effect, the rainy season starts still in the autumn (October) and reaches its peak in winter (January and February) before slowly subsiding in the spring (March and April) and therefore comes to an end anyway in the first week of May. Conversely, the historical and the forecasting data have almost no rainfall in the summer (June to September), which proves the continued dry summer climate in the area of study.

Source: Based on the results of the stacking rainfall forecasting model for the period (2025-2034) in Nineveh Governorate.

On a monthly basis, the forecast results indicate a relative increase in winter rainfall at a number of stations compared with historical values, particularly at the Rabia, Sinjar, Tal Afar, and Tal Abtah stations, which reflects a possible increase in the intensity of the winter season in the future. A slight extension of spring rains was also observed during April and May at some stations, which may indicate a limited lengthening of the rainy season compared with historical records. In contrast, other stations, such as Mosul and Al-Baaj, showed stability or a slight decrease in autumn rainfall, reflecting greater fluctuations at the beginning of the rainy season.

Comparing Figure 3 and the annual rainfall values between the predicted and historical data, one can find that a spatial difference exists between the stations. The values at Rabia, Sinjar, Tal Afar, and Tal Abtah were significantly higher than to the historical values. The Rabia station annual total increased from the historical 366 mm to a forecast value of more than 460 mm and Sinjar increased from the historical 346 mm to a forecast 394 mm, while Tal Afar increased from 319 mm to a forecast 393 mm. This increase is a relative positive variation of the future rainfall resources in the north and the western areas of the governorate. Other stations, such as Mosul, Makhmur, and Baaj, however, exhibited different trends with the stations being relatively stable or recording a decline in the total rainfall. At Mosul station, the annual total decreased to a projected under 290 mm /a by a traditional decrease of approximately 327 mm/a. At Makhmur station, the reduction was recorded at 365 to 317 mm Al-Baaj station was also close to being stationary with a slight lossThe difference indicates the dissimilarity in the spatial response to the future climate changes in the governorate. Having spatial comparisons meant that chances of increase in rainfall in the future are high in the northern and northern part of the Nineveh Governorate as compared to central and the south part which is likely to remain at a constant level or relatively lower amount of rainfall. The relevance of such a trend is the increasing spatial and temporal variability of rainfall that may directly apply to agricultural planning and water resource management in the study area.

Overall, the forecasting outcomes of the period (2025-2034) show that the rainfall regime in Nineveh Governorate will preserve the primary seasonal features of the onset and the conclusion of the rainy season, where quantitative changes of the monthly averages and annual totals are not identical in all of the stations. The findings indicate the relevance of integrating the use of sophisticated predictive models to aid in the comprehension of the future climate change and offer a concrete level of scientific foundation to inform decision making in areas of agriculture and water management.

Table 9 shows the results of the linear regression for rainfall forecasting at the Nineveh governorate, which reveals that the Mosul station records a weak downward trend of −1.97 mm/year with R² = 0.93 and P = 0.422, whereas the Baaj station shows almost complete rainfall stability with a slight increase of 0.44 mm/year with R² = 0.7 and P = 0.818. The Rabia station also records a limited upward trend of 2.23 mm/year, with R² = 0.71 and P = 0.416. The Tal Afar station shows an increase of 2.34 mm/year, with R² = 0.82 and P = 0.494, whereas the values at the Sinjar station increase to 3.20 mm/year, with R² = 0.64 and P = 0.348. The Makhmur station records an increase of 3.29 mm/year, with R² = 0.85 and P = 0.240. The Tel Abtah station recorded the highest annual change rate of 4.39 mm/year (R² = 0.91), which was close to statistical significance (P = 0.087), indicating that future rainfall changes in Nineveh Governorate are dominated by climatic fluctuations, with no significant long-term trend.

Source: Based on the results of the stacking rainfall forecasting model for the period (2025-2034) in Nineveh Governorate.

Table 10 shows the standard deviation values ​​for the expected rainfall during the period (2025- 2034). This value indicates an average spatial heterogeneity in rainfall volumes at stations within the Nineveh governorate, with values between 21.6 mm and 27.3 mm. This comparatively small area depicts a certain level of spatial homogeneity in the temporal variation pattern of precipitation. The differences in space exist and are associated with the geographical position of stations and topography of the area. The highest standard deviations were recorded at our stations Rabia (27.3 mm) and Sinjar (26.9 mm), which show that the north and northwest of the study area are more significantly exposed to the action of Mediterranean low-pressure systems and to variations in their routes and intensity. Conversely, expectant stability in the expected level of rainfall in the western and southern regions of the study area is observed by the fact that the standard deviation values of the Tal Afar (21.6 mm), the Al-Baaj (21.7 mm), and Makhmour (21.7 mm) stations are the lowest.

Source: Based on the results of the stacking rainfall forecasting model for the period (2025-2034) in Nineveh Governorate.

However, the Tel Abtah (22.6 mm) and Mosul (22.8 mm) stations recorded average values ​​of standard deviation, indicating a gradual shift in rainfall characteristics between the high-variability areas in the north and the more stable study area in the south.

In general, these results confirm that the spatial variation in rainfall in Nineveh Governorate will remain relatively limited in the future, with clear spatial differences emerging that reflect the role of geographical location and influential weather systems, which is consistent with the results of the analysis of annual and monthly rainfall totals at stations in the study area (Al Azawi, 2017).

Figure 4 shows the annual changes in the standard deviation coefficient of rainfall at all the study stations for the period (2025–2034), where the highest values ​​were recorded during the years 2026 and 2029, especially at the Mosul, Tal Afar, Makhmour and Sinjar stations, indicating high variability in rainfall during those years. In contrast, other years, such as 2028 and 2031, presented relatively lower values, reflecting a degree of relative stability in rainfall, although spatial variation between stations continued.

Figure 5 shows monthly changes in the standard deviation coefficient of rainfall at the study stations, with the highest values being recorded during the winter months, which are the months of January and February, especially at the Tal Afar, Mosul, Al-Baaj, and Tal Abtah stations, meaning that the rainfall is highly variable during the rainy seasons. Conversely, rainfall levels were significantly reduced at all study sites, especially in July, indicating drought conditions and near-nonuse of rainfall throughout summer (June-August). The average rainfall levels in the transitional seasons, i.e. spring and autumn, indicate low variations between relative stability and variability of rainfall.

Figure 5. Seasonal Changes in Monthly Rainfall for Each Study Station for the Period 2025–2034.

These spatial variability patterns also reflect the inherent uncertainty associated with rainfall prediction in semi-arid environments. The variability between stations, combined with the sensitivity of individual machine learning models to climatic fluctuations, indicates that model uncertainty cannot be neglected. However, the implementation of the weighted stacking ensemble approach helped reduce predictive uncertainty by integrating the strengths of individual models and minimizing their independent errors. This integration enhances the robustness and stability of the forecasting framework, providing more reliable projections under spatially heterogeneous climatic conditions.

The spatial distribution maps of rainfall in the study area Figure 6 illustrate significant spatial and temporal differences in rainfall amounts among the four climatic cycles, reflecting the widespread influence of climatic factors and environmental changes in shaping rainfall patterns at the study area level.

For example, during the period of 1994–2003, the highest rainfall amounts were concentrated in the northern and northwestern parts of the governorate, particularly in the areas of Rabia, Sinjar and Al-Baaj, where high values exceed (330–360 mm annually) prevailed, whereas the southern and southeastern regions, such as Tel Abtah and Makhmour, recorded relatively lower amounts, which indicates a decreasing rainfall gradient towards the south. During the period of 2004–2013, a relative decline in the intensity of rainfall concentration was observed, with the continued dominance of average values (280–320 mm) over most parts of the governorate, in contrast to the decrease in areas with high rainfall and their confinement to limited areas in the north, which reflects the beginning of a change in the general rainfall pattern.

In the period of 2014–2024, clearer features of rainfall irregularity emerged as the ranges of medium and low rainfall expanded, and the areas with high values decreased, with the relative concentration of rainfall remaining in the north and northwest, especially in Sinjar and Rabia, compared with a noticeable decrease in the central and southern regions. This trend is expected to persist in the period 2025-2034, according to the forecast map, as evidenced by the dominance of average and low rainfall across most of Nineveh Governorate, with regions experiencing heavy rainfall decreasing and concentrated in thin strips in the northern governorate. When the spatial distribution of rainfall in previous climatic periods Figure 6 is compared with the expected distribution for the period (2025-2034) Figure 7, it becomes clear that the general spatial pattern of precipitation continues in Nineveh Governorate, where the northern and northwestern regions retain the most abundant rainfall. However, the predictive results indicate quantitative changes in rainfall intensity, characterized by a shrinking range of high values and an expansion of medium and low values, particularly in the central and southern regions.

The spatial gradient from north to south continues, albeit with a relative decrease in intensity, reflecting increasing annual climatic variability. These data suggest that future changes represent an extension of the historical distribution with a relative redistribution of values, influenced by prevailing climatic factors. Overall, the results reflect a trend toward greater spatial variability and a decline in areas with high rainfall, reinforcing the prevalence of semiarid conditions and increasing pressure on water resources and agricultural stability.

Such outcomes suggest the reinforcement of arid and semiarid environmental features, since the significant increase in rainfall variability underscores the growing importance of climate change in reorganizing the spatial distribution of rainfall in the study area. Overall, rainfall trends across the climatic periods showed a downward shift, and variation and heterogeneity in spatial distribution were evident across the periods of investigation, as manifested by a gradual shift toward lower rainfall and a gradual reduction in high-rainfall areas. This rainfall pattern also demonstrates structural climatic change in the dominant hydroclimatic system, which implies further pressure on surface and groundwater resources and an increasing threat to the sustainability of agriculture and food security in Nineveh Governorate over the decades (Zelenakova et al., 2022).

The combined results of the Mann–Kendall trend analysis and the correlation assessments (Pearson and Spearman) reveal that rainfall variability in Nineveh Governorate during the period (1994–2024) is primarily governed by interannual climatic oscillations rather than statistically robust long-term monotonic trends. Although certain stations exhibited weak positive or negative Sen’s slope values, none reached a level sufficient to indicate structural climatic shifts.

The correlation structure indicates a tightly coupled thermo–humidity–pressure system, in which rainfall is strongly and negatively associated with temperature and solar radiation, and positively associated with relative humidity and atmospheric pressure. The near-perfect intercorrelation among thermal variables confirms multicollinearity, which was carefully addressed during model construction to ensure predictive stability.

Importantly, the absence of statistically significant long-term rainfall trends, combined with strong internal climatic coupling, suggests that rainfall forecasting performance depends more on dynamic inter-variable interactions than on persistent directional change. Therefore, the proposed modeling framework integrates statistically validated trend detection, robust dependence structure assessment, and cautious variable selection to enhance interpretability, reduce structural uncertainty, and improve forecasting reliability under semi-arid climatic conditions.