Deep Learning-Based Multi-Source Precipitation Forecasting in Arid Regions Using Different Optimizations: A Case Study from Konya, Turkey

Abstract

Accurate precipitation forecasting plays a crucial role in sustainable water resource management, especially in arid regions like Konya, one of Turkey’s driest areas. Reliable forecasts support effective water budgeting, agricultural planning, and climate adaptation efforts in the region. This study investigates the performance of different deep learning training algorithms in forecasting monthly precipitation using Long Short-Term Memory (LSTM) networks, a method tailored for time-series prediction. A comprehensive dataset comprising 39 years (1984–2022) of precipitation records was utilized, obtained from the Turkish State Meteorological Service (MGM) as ground-based observations and from NASA’s POWER database as remote sensing data, and was split into 80% for training and 20% for testing. A comparative analysis of three widely used optimization algorithms, Adaptive Moment Estimation (ADAM), Root Mean Square Propagation (RMSProp), and Stochastic Gradient Descent with Momentum (SGDM), revealed that ADAM consistently outperformed the others in forecasting accuracy. Model performance was evaluated with statistical metrics, and the LSTM-ADAM combination achieved the best results. In the final phase, cross-validation was applied using MGM and NASA data sources in a crosswise manner to test model generalizability and data source independence. The best performance was observed when the model was trained with MGM data and tested with NASA data, achieving a remarkably low RMSE of 3.62 mm, MAE of 2.93 mm, R² of 0.9966, and NSE of 0.9686. When trained with NASA data and tested with MGM data, the model still demonstrated strong performance, with an RMSE of 4.48 mm, MAE of 3.22 mm, R² of 0.9921, and NSE of 0.9678. These results demonstrate that satellite and ground-based data can be used interchangeably under suitable conditions, while also confirming the superiority of the ADAM optimizer in LSTM-based precipitation forecasting.

1. Introduction

Civil engineering plays a central role in designing and managing the environment. Within this discipline, hydraulic engineering has increasingly benefited from the integration of machine learning techniques, which provide powerful tools for addressing complex challenges such as water resources management, flood and drought forecasting, and sustainable planning under climate change. By uncovering hidden patterns in large hydrometeorological datasets, machine learning approaches contribute to improved decision-making and operational efficiency in water-related projects. Historically, water has been a fundamental driver of human civilization, with early settlements flourishing near rivers and lakes due to their importance for agriculture, trade, and industry. Today, the ability to accurately model and forecast precipitation and hydrological processes remains critical for sustaining human and ecological systems [1]. This necessity underlines the central role of water as a vital resource across multiple sectors of society. As an essential resource for both human life and ecosystems, water is critically important across various sectors, including agriculture, industry, and energy production. Among these, the agricultural sector is particularly dependent on water availability, necessitating the implementation of effective water management strategies. In this context, accurate modeling and forecasting of precipitation data are of paramount importance for the sustainable management of water resources [2]. Precipitation data serve as a key component in water resource planning, and their precise estimation is crucial for ensuring the efficient and sustainable use of water, particularly in water-intensive sectors such as agriculture [3].

In recent years, artificial intelligence and deep learning techniques have offered significant advantages in the analysis and forecasting of meteorological data. Owing to their ability to learn complex patterns from large datasets, deep learning models have demonstrated superior accuracy compared to traditional approaches. For instance, Jihoon Ko and colleagues highlight the application of deep learning methods in weather-related tasks such as precipitation forecasting, emphasizing their advantages over conventional techniques [4]. When integrated with data derived from meteorological observations as well as remote sensing technologies, these methods enable faster, more accurate, and more reliable predictions. In the literature, studies on modeling precipitation data have generally been limited to traditional statistical methods. However, in recent years, the use of artificial intelligence and machine learning techniques has increased in this field. For example, Shilpa Manandhar and colleagues developed a data-driven approach for precipitation detection using meteorological sensor data. In this study, the effects of certain atmospheric parameters on precipitation probability were investigated and a correct detection rate of 87.4% was obtained [5]. Similarly, Waqas et al., achieved high accuracy rates in daily rainfall prediction with a hybrid deep learning method they developed in Thailand and provided significant improvements compared to traditional models [6]. In another study, a deep learning model applied to rainfall forecasting in Ratnapura, Sri Lanka, demonstrated strong capability in predicting extreme rainfall events, achieving an RMSE < 5 mm and an MAE < 3 mm [7]. Such studies show how effective remote sensing data, artificial intelligence techniques, and especially LSTM can be in precipitation prediction.

The POWER dataset, a platform developed by NASA utilizing remote sensing techniques, provides a valuable resource for the analysis of monthly, climatic, and annual seasonal data. These data contribute significantly to a better understanding of climatic variables, particularly when compared with observational records. For instance, Abdul Hasib Halimi and colleagues compared NASA POWER reanalysis data with observations from meteorological stations located in the Mediterranean and continental regions of Turkey, and reported that the dataset demonstrated high accuracy in estimating temperature and relative humidity [8]. Tan et al. conducted a thorough study of gridded climate products, notably NASA POWER and ERA5-Land, in a tropical river basin, highlighting both their promise and limits in capturing precipitation dynamics and extremes. Their findings emphasize the necessity of combining sophisticated data analysis approaches with remote sensing and reanalysis products to improve hydrological modeling and severe event forecasting in unmonitored or data-poor locations [9]. A thorough review of the literature was carried out utilizing the keywords “precipitation, AND LSTM” in the Web of Science database. The assessments conducted in 2025 showed that, including the relevant keywords, there were 985 academic studies in total. VOSviewer V1.6.20 software was used to connect the studies using a restricted set of keywords [10,11]; Figure 1 provides a visual representation of the connections among the identified keywords.

In recent years, there has been a marked shift in researchers’ focus on advanced data-driven methods in hydrology. As shown in Figure 1, keywords such as deep learning, LSTM, precipitation, and climate change occupy central positions in the research landscape. This trend reflects a growing preference for deep learning, LSTM, and various machine learning approaches that have been widely adopted due to their superior predictive capabilities compared to traditional statistical methods [12,13,14,15].

In this study, monthly precipitation data from Konya, one of Turkey’s driest regions, were collected for the period 1984–2022 from both ground-based meteorological stations and NASA’s POWER database. The main objective was to assess the capability of deep learning methods, specifically Long Short-Term Memory (LSTM) networks, in forecasting monthly precipitation under arid conditions. For model development, the dataset was divided into training (80%) and testing (20%) subsets. Three widely used optimization algorithms—Adaptive Moment Estimation (ADAM), Root Mean Square Propagation (RMSProp), and Stochastic Gradient Descent with Momentum (SGDM)—were compared using standard statistical metrics to evaluate forecasting accuracy and model robustness.

To further examine generalizability, a cross-validation framework was implemented by training on one data source (MGM ground-based or NASA/POWER satellite) and testing on the other. This approach enabled the evaluation of consistency and reliability across independent datasets, thereby assessing their potential interchangeability.

The novelty of this work lies in:

❖

Integrating multi-source datasets (MGM and NASA/POWER) for precipitation forecasting in a semi-arid region.

❖

Benchmarking multiple optimization algorithms within the LSTM framework.

❖

Conducting cross-validation between independent data sources to test model robustness and transferability.

By addressing these aspects, the study clarifies both the practical potential and the limitations of deep learning models for operational water management and climate adaptation in arid regions.

2. Materials and Methods

2.1. Study Area

The Konya Closed Basin (KCB) is one of the most significant endorheic basins located in the central part of Turkey’s Central Anatolia Region (Figure 2). Covering an area of approximately 50,000 km², the basin is of critical importance both hydrologically and ecologically. Surrounded by mountains, this closed basin possesses a unique drainage structure that prevents natural outflow to the sea, resulting in the retention of water within the basin. The region is characterized by a semi-arid continental climate, with limited and irregular annual precipitation. Agriculture constitutes the dominant land use in the area, with water demands largely met through groundwater and surface water resources [16,17]. However, due to increasing anthropogenic pressures and climatic variability, significant declines in groundwater levels and shrinkage of wetland areas have been observed.

The management of water resources in the KCB is vital for sustaining agricultural productivity in the region. Agricultural practices have led to excessive exploitation of water resources, contributing to considerable declines in groundwater levels. For instance, a study conducted by Bayari et al. (2009) reported an average annual decline of approximately 1 meter in groundwater levels across the KCB [18]. This situation creates challenges in securing the water necessary for agricultural irrigation and poses a threat to the ecological balance of the region. Moreover, the impacts of climate change in the KCB further complicate water resource management. A meteorological drought analysis conducted by Sarış and Gedik (2021) revealed an increasing trend in drought conditions within the KCB, which has had adverse effects on agricultural productivity. The study examined drought severity using data collected from 11 meteorological stations between 1930 and 2019. The findings underscore the urgent need for effective and sustainable water resource management strategies to address the growing water scarcity in the region [19]. The management of water resources in the KCB is also directly linked to the monitoring and assessment of groundwater levels. In a study conducted by [20], the effects of groundwater level variations on the positions of continuously operating GNSS stations were investigated. The study indicated that the decline in groundwater levels was primarily attributed to drought conditions resulting from global climate change and excessive water usage. The management of water resources in the region is also of great importance for agricultural planning [20]. A study conducted by Torun and Çakmak [21] evaluated agricultural water efficiency in the KCB [21]. The research analyzed annual irrigation water distribution volumes and efficiency indicators, emphasizing that such efficiency assessments are crucial for enhancing agricultural productivity and ensuring more effective utilization of water resources. Another key issue related to water management in the KCB is the monitoring of groundwater levels and their impact on agricultural productivity. In a study, the effects of groundwater level fluctuations on ground pressure were examined. The findings revealed that declines in groundwater levels could disrupt ground stability, potentially leading to adverse consequences in agricultural areas [22].

Figure 2 illustrates the geographical location of the KCB. The basin boundaries are shown with red lines, while major rivers and lakes are indicated in blue. Elevation differences are represented using a color gradient (green to white), providing topographic context. The average elevation of the basin is approximately 1000 m.

2.2. Data

The data used in this study consist of monthly total precipitation records for the province of Konya (37.8667° N, 32.4833° E), located within the KCB of Turkey, covering the period from 1984 to 2022. The data were obtained from two primary sources: the first source comprises ground-based observations provided by MGM (https://www.mgm.gov.tr/ (accessed on 17 October 2025) ), the official institution responsible for conducting meteorological observations and providing weather forecasting services across Turkey. MGM has been delivering reliable measurements nationwide for many decades. The second source involves remote sensing data retrieved from NASA’s POWER (https://power.larc.nasa.gov/ (accessed on 17 June 2025)) platform. These remote sensing data are derived from the MERRA-2 (Modern-Era Retrospective Analysis for Research and Applications, Version 2) dataset, which includes a wide range of climatic parameters such as solar radiation (ALLSKY_SFC_SW_DWN), total bias-corrected precipitation (PRECTOTCORR), relative humidity at 2 m (RH2M), temperature at 2 m (T2M), and wind speed at 2 m (WS2M) [23,24].

NASA data are particularly preferred for large-scale analyses due to their ability to provide consistent and continuous observations at a global scale. In this study, both ground-based data from the MGM and satellite-based data from NASA were evaluated separately. A correlation coefficient of 0.888 was found between the two datasets, indicating a strong relationship. To further harmonize the satellite-derived precipitation values with ground-based MGM observations, a bias correction function of (1.58x + 24.90) was applied. Figure 3 below visually presents the temporal variation in precipitation data, and statistical information about the data is presented in

Table 1. Statistical information of data.

Parameter	Observed MGM	Observed NASA
Mean	26.77	26.77
Median	20.9	21.63
Standard Deviation	24.04	21.35
Max	124.00	145.02
Min	0.00	0.00

.

In Figure 3, the blue line represents ground-based MGM observations, while the black line corresponds to NASA/POWER satellite data. Both series are shown in millimeters (mm). The plot highlights seasonal and interannual variability as well as the overall consistency between the two datasets.

2.3. LSTM

Deep learning plays a significant role, particularly in forecasting tasks involving time series data. In this context, the LSTM architecture is frequently utilized due to its ability to model temporal dependencies and handle long-range patterns within sequential data. Originally proposed by Hochreiter and Schmidhuber (1997) as a solution to the vanishing gradient problem in traditional Recurrent Neural Networks (RNNs), LSTM has since become one of the most widely adopted deep learning techniques for time series analysis across various disciplines, including hydrology, climate science, and environmental modeling [25]. LSTM was specifically developed to address the limitations of traditional RNNs [26]. Its primary advantage lies in its capacity to learn long-term dependencies. LSTM cells are capable of retaining information over extended time periods, thereby allowing the model to capture the complex structures inherent in time series data. For instance, in one study employing the LSTM architecture, the model achieved high accuracy in time series forecasting tasks [27]. An LSTM unit consists of a memory cell equipped with a forget gate (f_t), an input gate (i_t), and an output gate (o_t), which collectively govern the flow of information. At each time step t, the hidden state h_t is updated using the current input x_t, the previous hidden state h_t−1, and the previous cell state c_t−1. Figure 4 displays both the fundamental architecture of LSTM and the internal structure of its memory unit.

This structure enables LSTM to learn complex relationships within time series data [28]. A recent study introduced a model that illustrates the application of LSTM networks in multi-step time series forecasting. By utilizing the multi-layered architecture of LSTM, the model is designed to effectively capture the underlying patterns and characteristics of time series data [29]. Such structures enhance the overall performance of LSTM and contribute to improved forecasting accuracy. Moreover, the use of LSTM has also been observed in meteorological parameter prediction tasks. In one study, an LSTM-based model outperformed traditional methods by achieving significantly higher accuracy rates [24]. These findings suggest that LSTM can contribute not only to addressing the critical issue of water scarcity in the region but also to exploiting its potential in energy forecasting. The forward pass process of LSTM across time steps can be expressed through the following vectorized formulas [30]:

Forget gate:

$${\mathrm{f}}_{\mathrm{t}}=\mathsf{\sigma }\left({\mathrm{W}}_{\mathrm{h}\mathrm{f}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{x}\mathrm{f}}{\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{f}}\right)$$

(1)

Input gate:

$${\mathrm{i}}_{\mathrm{t}}=\mathsf{\sigma }\left({\mathrm{W}}_{\mathrm{h}\mathrm{i}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{w}}_{\mathrm{x}\mathrm{i}}{\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{i}}\right)$$

(2)

Output gate:

$${\mathrm{O}}_{\mathrm{t}}=\mathsf{\sigma }\left({\mathrm{W}}_{\mathrm{h}\mathrm{o}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{x}\mathrm{o}}{\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{o}}\right)$$

(3)

New value:

$${\stackrel{\sim }{\mathrm{C}}}_{\mathrm{t}}=\tanh \left({\mathrm{W}}_{\mathrm{h}\mathrm{c}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{x}\mathrm{o}}{\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{c}}\right)$$

(4)

Cell state:

$${\mathrm{C}}_{\mathrm{t}}={\mathrm{f}}_{\mathrm{t}}\otimes {\mathrm{C}}_{\mathrm{t}-1}+{\mathrm{i}}_{\mathrm{t}}\otimes {\stackrel{\sim }{\mathrm{C}}}_{\mathrm{t}}$$

(5)

Hidden state:

$${\mathrm{h}}_{\mathrm{t}}={\mathrm{O}}_{\mathrm{t}}\otimes \tanh \left({\mathrm{C}}_{\mathrm{t}}\right)$$

(6)

The parameters of the gates include weights (W_i, W_f, W_c, W_o) and biases (b_i, b_f, b_c, b_o) corresponding to the input (x_t) and prior hidden state (h_t−1). The cell state flowing through the input gate ranges from −1 to 1. The memory cell has two states: present (C_t) and previous (C_t−1). “tanh” is the hyperbolic tangent activation function, while σ is the sigmoid activation function. ⊗ represents the convolutional operation (Hadamard product). The cell’s output (h_t) is determined by its current state and output gate [24].

At the end of the methodological framework, the overall workflow of the study is summarized and visually presented in Figure 5 to provide a clearer understanding of the sequential steps applied.

In Figure 5, To ensure the robustness of the model, the dataset was initially divided internally into 80% for training and 20% for testing. In the second stage, a cross-validation approach was applied to further evaluate the model’s generalization ability in more detail.

3. Results

In this study, LSTM was implemented to forecast monthly precipitation in Konya. The dataset, covering the period 1984–2022, was obtained from both MGM and NASA POWER sources. The data were split into training (80%) and testing (20%) sets. Three optimization algorithms—ADAM, RMSProp, and SGDM—were employed during the training phase to assess their effect on model performance. Evaluation was conducted using RMSE, MAE, R² and NSE metrics. To further validate the model’s generalizability, a cross-validation approach was applied using two configurations: one where NASA POWER data were used for training and MGM data for testing, and another with the reverse setup.

In the modeling process, LSTM architecture was employed for monthly precipitation prediction using a univariate time-series dataset. The model was trained and tested with different configurations through a nested loop structure, where the number of training epochs (m) varied from 100 to 300, and the number of hidden units in the LSTM layer (n) ranged from 10 to 30. Input and output variables were standardized using z-score normalization to improve model convergence. The LSTM model architecture consisted of a “sequenceInputLayer”, a single “lstmLayer” with n hidden units, followed by a “fullyConnectedLayer” and a “regressionLayer”. The learning process was configured with a maximum number of epochs (MaxEpochs) set to m, an initial learning rate of 0.05, and a piecewise learning rate schedule that reduced the rate by a factor of 0.2 every 125 epochs. A gradient threshold of 1 was applied to prevent gradient explosion, as suggested in [31]. Models are evaluated using the commonly recommended criteria in the literature [32,33]. The analysis used comparison measures such as RMSE, MAE, R² and NSE [34,35]. The formulas for each comparison criterion are provided below.

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{Y}}_{\mathrm{e}}-{\mathrm{Y}}_{\mathrm{o}}\right)}^2}$$

(7)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\left|{\mathrm{Y}}_{\mathrm{e}}-{\mathrm{Y}}_{\mathrm{o}}\right|$$

(8)

$${\mathrm{R}}^2={\left({\displaystyle \frac{\mathrm{N}\left(\sum {\mathrm{Y}}_{\mathrm{o}}{\mathrm{Y}}_{\mathrm{e}}\right)-\left(\sum {\mathrm{Y}}_{\mathrm{o}}\right)\left(\sum {\mathrm{Y}}_{\mathrm{e}}\right)}{\sqrt{\left(\mathrm{N}\sum {\mathrm{Y}}_{\mathrm{o}}^2-{\left(\sum {\mathrm{Y}}_{\mathrm{o}}\right)}^2\right)\left(\mathrm{N}\sum {\mathrm{Y}}_{\mathrm{e}}^2-{\left(\sum {\mathrm{Y}}_{\mathrm{e}}\right)}^2\right)}}}\right)}^2$$

(9)

$$\mathrm{N}\mathrm{S}\mathrm{E}=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{Y}}_{\mathrm{o}}-{\mathrm{Y}}_{\mathrm{e}}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{Y}}_{\mathrm{o}}-{\overline{\mathrm{Y}}}_{\mathrm{o}}\right)}^2}}$$

(10)

In these equations, N represents the number of data, Y_e represents the model-predicted data, and Y_o represents the observed data. Since RMSE and MSE are among the primary assessment criteria utilized in this study and both measure error magnitudes, their values are given in millimeters (mm). Higher error values indicate that the model’s predictions differ greatly from the observed data, implying poor performance. In contrast, lower error levels indicate that the model closely approximates the real data, meaning more accuracy and dependability. The coefficient of determination (R²) scales from 0 to 1, with higher values suggesting a greater correlation between predicted and observed values, resulting in a more accurate and consistent model. The comprehensive findings of the analysis are shown in

Table 2. Modeling results of different parameters and optimization algorithms for precipitation forecasting using MGM data.

Dataset	Algorithm	Epochs	Hidden Layers	Train				Test
				Train RMSE	Train MAE	Train R2	Train NSE	Test RMSE	Test MAE	Test R2	Test NSE
MGM	ADAM	100	10	1.216	1.089	0.999	0.997	3.375	1.851	0.989	0.982
		100	20	1.266	1.090	0.999	0.997	3.952	2.150	0.982	0.975
		100	30	0.907	0.681	0.999	0.999	4.080	2.798	0.990	0.973
		200	10	0.502	0.395	0.999	0.999	3.642	2.440	0.991	0.979
		200	20	0.263	0.204	0.999	0.999	4.306	2.857	0.985	0.970
		200	30	0.334	0.276	0.999	0.999	4.016	2.993	0.991	0.974
		300	10	0.348	0.259	0.999	0.999	3.765	2.695	0.991	0.977
		300	20	0.174	0.138	0.999	0.999	4.218	2.802	0.985	0.971
		300	30	0.180	0.135	0.999	0.999	3.874	2.782	0.991	0.976
	RMSProp	100	10	3.387	2.445	0.994	0.980	6.239	4.364	0.989	0.938
		100	20	3.063	2.421	0.990	0.983	5.070	3.400	0.968	0.959
		100	30	4.352	3.762	0.994	0.966	8.284	6.955	0.984	0.890
		200	10	1.016	0.784	0.999	0.998	4.147	3.240	0.993	0.972
		200	20	1.293	1.014	0.999	0.997	5.312	3.608	0.985	0.955
		200	30	0.738	0.586	0.999	0.999	4.897	3.707	0.990	0.962
		300	10	0.647	0.465	0.999	0.999	3.520	2.588	0.995	0.980
		300	20	0.511	0.354	0.999	0.999	4.400	2.956	0.986	0.969
		300	30	0.386	0.302	0.999	0.999	4.679	3.253	0.987	0.965
	SGDM	100	10	3.917	3.516	0.995	0.973	7.628	6.175	0.970	0.907
		100	20	6.945	6.789	0.996	0.914	6.543	5.216	0.965	0.931
		100	30	3.296	3.123	0.998	0.981	4.512	2.185	0.971	0.967
		200	10	0.979	0.733	0.998	0.998	4.539	2.608	0.979	0.967
		200	20	1.558	1.420	0.999	0.996	5.291	4.040	0.981	0.955
		200	30	2.392	2.253	0.999	0.990	6.132	4.981	0.978	0.940
		300	10	0.748	0.542	0.999	0.999	4.514	2.676	0.981	0.967
		300	20	0.682	0.505	0.999	0.999	4.552	3.006	0.981	0.967
		300	30	0.694	0.505	0.999	0.999	4.588	2.808	0.978	0.966

(for MGM data) and

Table 3. Modeling results of different parameters and optimization algorithms for precipitation forecasting using NASA/POWER data.

Dataset	Algorithm	Epochs	Hidden Layers	Train				Test
				Train RMSE	Train MAE	Train R2	Train NSE	Test RMSE	Test MAE	Test R2	Test NSE
NASA/POWER	ADAM	100	10	1.811	1.607	0.999	0.993	2.584	2.185	0.994	0.984
		100	20	1.357	1.259	0.999	0.996	2.503	2.044	0.995	0.985
		100	30	1.104	0.958	0.999	0.997	1.926	1.530	0.997	0.991
		200	10	0.652	0.526	0.999	0.999	1.978	1.165	0.995	0.991
		200	20	0.282	0.237	0.999	0.999	1.783	1.099	0.996	0.992
		200	30	0.151	0.113	0.999	0.999	1.549	1.062	0.998	0.994
		300	10	0.292	0.223	0.999	0.999	1.848	1.121	0.995	0.992
		300	20	0.144	0.109	0.999	0.999	1.781	1.159	0.997	0.992
		300	30	0.110	0.084	0.999	0.999	1.521	1.061	0.998	0.995
	RMSProp	100	10	3.517	2.601	0.984	0.974	4.433	3.242	0.980	0.953
		100	20	4.173	3.206	0.984	0.963	4.360	3.630	0.979	0.955
		100	30	4.654	3.630	0.981	0.954	5.768	4.669	0.963	0.921
		200	10	1.188	0.876	0.999	0.997	1.953	1.405	0.995	0.991
		200	20	0.984	0.810	0.999	0.998	1.902	1.239	0.994	0.991
		200	30	0.888	0.700	0.999	0.998	1.939	1.289	0.991	0.991
		300	10	0.535	0.404	0.999	0.999	1.488	1.124	0.996	0.995
		300	20	0.397	0.311	0.999	0.999	1.627	1.211	0.996	0.994
		300	30	0.382	0.305	0.999	0.999	2.014	1.484	0.992	0.990
	SGDM	100	10	2.577	2.279	0.996	0.986	3.020	2.178	0.989	0.978
		100	20	2.620	2.365	0.996	0.985	3.927	2.478	0.976	0.963
		100	30	3.032	2.821	0.997	0.980	4.016	2.867	0.981	0.962
		200	10	2.222	1.987	0.998	0.989	2.946	2.521	0.993	0.979
		200	20	1.731	1.539	0.998	0.994	3.267	2.297	0.985	0.975
		200	30	2.161	2.003	0.998	0.990	3.301	2.676	0.987	0.974
		300	10	0.860	0.572	0.998	0.998	1.897	1.333	0.994	0.991
		300	20	0.724	0.492	0.999	0.999	2.598	1.386	0.986	0.984
		300	30	1.109	0.913	0.999	0.997	2.617	1.749	0.989	0.984

(for NASA POWER data).

The modeling findings in Table 2 show how epoch count, and hidden layer size affect the performance of LSTM models trained on MGM data using three distinct optimization algorithms: ADAM, RMSProp, and SGDM. The ADAM optimizer produced the best training results, with decreasing RMSE and MAE values as the number of epochs grew, and R² and NSE values reaching 1.000. However, while training errors dropped with increased complexity (more epochs and hidden units), test performance fluctuated somewhat, showing that extremely complex setups may result in overfitting. RMSProp, on the other hand, had significantly more training errors, particularly at lower epoch counts, but produced the most consistent and accurate test results in simpler designs. The setup with 300 epochs and 10 hidden units had the lowest test RMSE (3.520) and greatest R² (0.995), indicating good generalization ability. SGDM performed poorly in early iterations, especially with 100 epochs, but improved dramatically with larger epoch counts, reaching test R² values of 0.97 in 200- and 300-epoch settings. However, the test RMSE values for SGDM were somewhat higher than those for ADAM and RMSProp.

The modeling results (Table 3) using NASA/POWER data reveal the performance differences among the ADAM, RMSProp, and SGDM optimization algorithms under varying LSTM configurations. ADAM once again outperformed the other algorithms, especially as the number of epochs and hidden layer units increased. The best results were achieved with 300 epochs and 30 hidden units, where training RMSE dropped to 0.110 and test RMSE to 1.521, with R² = 0.998 and NSE = 0.995, indicating exceptional accuracy and generalization capacity. ADAM consistently maintained low training errors across all configurations, and its test performance improved with deeper structures, suggesting that it can handle model complexity well without overfitting when trained sufficiently. RMSProp showed a clear improvement as training epochs increased, particularly when using simpler architecture. While it performed poorly at 100 epochs—with high training and test RMSE values—it provided competitive results at 300 epochs, notably achieving test RMSE = 1.488 and test R² = 0.996 with 10 hidden units. This suggests that RMSProp is highly effective when paired with extended training in lower-complexity models, although its performance slightly fluctuated with deeper networks. SGDM, while generally less accurate than ADAM and RMSProp, still demonstrated solid improvements at higher epoch levels. At 300 epochs and 10 hidden units, SGDM produced training RMSE = 0.860 and test RMSE = 1.897, with R² = 0.994, which is on par with RMSProp. However, deeper SGDM models (with 20–30 hidden units) resulted in slightly higher test errors and lower R², indicating a sensitivity to network complexity and the need for tuning. The findings show that ADAM proved to be the most effective optimizer for both training and testing on the MGM and NASA/POWER dataset, especially in deeper architectures and extended epochs. RMSProp delivered highly competitive test performance in shallow, well-trained models, while SGDM showed reliable but slightly less efficient behavior, requiring longer training to match the performance of the others. The scatter and time series plots corresponding to the best-performing methods are presented below in Figure 6 and Figure 7.

This study also evaluates the performance of LSTM under different optimization algorithms through Taylor diagrams, which visually illustrate forecasting accuracy and consistency with observed data. The diagrams are presented below in Figure 8 and Figure 9.

To further evaluate the distribution and variability of model predictions, a violin plot was employed. Violin plots combine box plot features with kernel density estimation, allowing a detailed visualization of prediction spread and central tendency. As shown in Figure 10 and Figure 11, LSTM-ADAM produces distributions that closely resemble the observed data and provide more reliable predictions.

Then, the test results of the applied models were statistically evaluated t-tests in order to assess the robustness and significance of differences between the measured and estimated precipitation values [36]. The tests were conducted at a 95% confidence level.

Table 4. Two-sample t-tests comparing the models during the test period.

Data	Method	t-Tests
		t	p
MGM	ADAM	0.4665	0.6413
	RMSProp	0.7292	0.4667
	SGDM	0.2179	0.8277
NASA	ADAM	0.1575	0.8750
	RMSProp	0.1143	0.9091
	SGDM	0.1536	0.8791

presents the t-test statistics.

In Table 4, t-tests revealed no statistically significant differences between the measured and modeled precipitation for any optimizer or data source (all p ≥ 0.467, α = 0.05). Accordingly, the LSTM-based estimates may be considered unbiased with respect to the observational mean during the test period, consistently across both MGM and NASA inputs. In this study, to develop the model in long-term precipitation forecasts and to test the data source independence, MGM ground-based data and NASA/POWER satellite data were used crosswise. With this cross-validation method applied in the last stage of the study, NASA data were used for training in the first scenario and MGM data for testing; in the second scenario, the opposite structure was adopted. This approach aimed to evaluate whether deep learning models (especially LSTM-ADAM) can provide consistent and reliable results with different data sources. The performance results obtained from these cross-validation scenarios are presented in

Table 5. Comparison of the best algorithm in predicting monthly total precipitation by using the cross data.

Dataset	Epochs	Hidden Layers	Train				Test
			RMSE	MAE	R2	NSE	RMSE	MAE	R2	NSE
MGM to NASA	100	10	1.2162	1.0885	0.999	0.9974	3.4302	2.6674	0.9956	0.9719
	100	20	1.2664	1.0895	0.999	0.9971	3.8078	2.9252	0.9925	0.9654
	100	30	0.9066	0.6808	0.999	0.9985	3.0756	2.492	0.9963	0.9774
	200	10	0.502	0.3952	0.999	0.9996	3.3564	2.6891	0.9965	0.9731
	200	20	0.2626	0.2044	0.999	0.9999	3.784	3.0416	0.9939	0.9658
	200	30	0.3337	0.2763	0.999	0.9998	3.6826	2.9927	0.9968	0.9676
	300	10	0.3475	0.2594	0.999	0.9998	3.4171	2.7842	0.9968	0.9721
	300	20	0.1741	0.1376	0.999	0.9999	3.8089	3.0673	0.9942	0.9653
	300	30	0.1797	0.1348	0.999	0.9999	3.6238	2.9344	0.9966	0.9686
NASA to MGM	100	10	1.8112	1.6068	0.999	0.9930	5.971	4.3221	0.9866	0.9428
	100	20	1.3571	1.2585	0.999	0.9960	5.6461	3.9396	0.9845	0.9488
	100	30	1.104	0.9576	0.999	0.9974	5.2329	3.7564	0.9909	0.9561
	200	10	0.6517	0.526	0.999	0.9991	4.3876	2.9649	0.9868	0.9691
	200	20	0.2819	0.2365	0.999	0.9998	4.5497	3.083	0.9865	0.9668
	200	30	0.1513	0.1127	0.999	0.9998	4.4567	3.2096	0.9919	0.9681
	300	10	0.2918	0.2226	0.999	0.9998	4.6575	3.2081	0.9878	0.9652
	300	20	0.1443	0.109	0.999	0.9998	4.6084	3.1196	0.9874	0.9659
	300	30	0.1101	0.0839	0.999	0.9998	4.4802	3.2224	0.9921	0.9678

, highlighting the robustness of the LSTM-ADAM algorithm across varying data sources.

The results in Table 5 show the performance of the LSTM model with different data sources in the cross-validation analysis performed between the MGM and NASA datasets. In the scenario trained with MGM data and tested with NASA data, the model generally exhibited lower error rates and higher accuracy. The best result obtained in the testing phase in this scenario was achieved when 300 epochs and 30 hidden layers were used; In this configuration, RMSE 3.62 mm, MAE 2.93 mm, R² 0.9966 and NSE 0.9686 values were obtained. On the other hand, it was observed that the error rates were relatively higher in the scenario trained with NASA data and tested with MGM data. In this second scenario, the most successful result was again obtained with 300 epochs and 30 hidden layers, and in the testing phase, RMSE 4.48 mm, MAE 3.22 mm, R² 0.9921 and NSE 0.9678 values were reached. Although the ADAM optimization algorithm showed superior performance in both scenarios, the overall evaluation showed that the model trained with MGM data and tested with NASA data stood out with lower error and higher accuracy. As a result, the most successful performance was obtained in training with MGM dataset, testing with NASA dataset, 300 epochs and 30 hidden layers configuration.

4. Discussion

In this study, the primary focus was to enhance the accuracy of precipitation forecasting, which plays a critical role in sustainable water resource management in Konya. LSTM networks were selected due to their proven effectiveness in modeling time series data, particularly when dealing with complex and dynamic environmental patterns. Previous research has consistently demonstrated the superior performance of LSTM models among deep learning techniques, making them a popular choice for hydrometeorological forecasting tasks [12]. In line with these findings, this study highlights the advantage of LSTM’s ability to learn from sequential historical data, which significantly contributes to improved forecasting precision. A comparative analysis of various deep learning training algorithms revealed that the ADAM optimizer consistently outperformed its counterparts in terms of predictive accuracy. This observation aligns with existing literature, which emphasizes ADAM’s ability to accelerate convergence while enhancing model accuracy [37]. The superior performance of ADAM in this study can be attributed to its adaptive learning rate mechanism, which provides stability and prevents overfitting when dealing with heterogeneous climatic time series, whereas RMSProp and SGDM are more sensitive to local minima. Accordingly, the use of ADAM in optimizing LSTM models is strongly supported by the results of this study. Model performance was rigorously evaluated using standard statistical metrics, including RMSE, MAE and the R² [38,39]. Another critical component of this study was the implementation of cross-validation to assess the generalizability of the model. This evaluation was conducted using data from two different sources: the MGM and NASA. By training the model using MGM (ground-based) data and testing it with NASA (remote-sensing) data, this study aimed to evaluate the model’s robustness across independent data sources. Although both datasets share the same spatial coordinate framework, the inclusion of remote sensing–based variables with broader spatial coverage helps to mitigate the potential impact of domain shift between ground-based observation stations and large-scale predictors. This integrated strategy enhances the model’s capacity to generalize under varying data distributions. The results demonstrated that the model maintained a high level of performance even when applied to previously unseen data, indicating strong robustness and data-source independence. This can be explained by the fact that MGM station data directly captures the local hydrometeorological dynamics of the Konya Basin (semi-arid continental climate, groundwater-driven feedbacks), which are sometimes smoothed in reanalysis or satellite-based datasets, thereby allowing LSTM to learn more localized precipitation variability. This finding suggests that, under appropriate conditions, remote sensing and ground-based observations can be used interchangeably for precipitation forecasting [40,41]. Moreover, the application of LSTM in weather forecasting holds considerable importance, particularly in the context of climate change adaptation and sustainable water resource management. The ability of LSTM networks to capture nonlinear temporal dependencies allows for the effective modeling of complex meteorological variables such as precipitation, temperature, humidity, and wind speed. This capability is particularly critical given the increasing variability and extremity of weather patterns driven by global climate change. Numerous studies have demonstrated the superior performance of LSTM-based models in forecasting weather-related parameters with higher accuracy compared to traditional statistical methods. For instance, LSTM has been successfully used to enhance short- and medium-term forecasts, which are vital for early warning systems, drought assessments, and precision agriculture applications. Additionally, improved weather prediction facilitated by LSTM contributes to better decision-making in water allocation, energy demand forecasting, and climate-resilient infrastructure planning [42,43,44,45,46,47,48]. Given these advantages, the integration of LSTM models into operational forecasting systems can play a transformative role in managing climate-related risks and ensuring the resilience of socio-ecological systems [13,49]. In this context, it is emphasized that LSTM models can play a critical role in addressing climate change and supporting sustainable water resource management. The effectiveness of LSTM networks in time series forecasting has been further evaluated through comparisons with various optimization algorithms, including ADAM, RMSProp, and SGDM. Among these, the ADAM optimizer consistently produced superior results. Notably, the model trained using MGM data and tested with NASA data achieved an RMSE of 3.62 mm, an MAE of 2.93 mm, an R² of 0.9966, and an NSE of 0.9686. These high-performance metrics provide compelling evidence that remote sensing data and ground-based observations can be used interchangeably under appropriate conditions. This is particularly significant in the context of integrated water management and climate resilience, where reliable data may be scarce or unevenly distributed. Moreover, studies exploring the generalizability of LSTM-based forecasting models across different data sources demonstrate the flexibility of such models. Their ability to learn from historical patterns and accurately predict future precipitation supports their practical potential in water resources planning, drought risk mitigation, and adaptive infrastructure development [23,50,51,52]. The findings of this study offer valuable insights into water management and climate resilience, particularly in arid regions such as Konya. The high accuracy achieved by LSTM-based models in precipitation forecasting enables more effective planning and management of water resources. Furthermore, the integration of remote sensing and ground-based data emerges as a critical strategy for enhancing the reliability of such forecasts.

5. Conclusions

In this study, the performance of LSTM-based deep learning models was evaluated using ground-based (MGM) and remotely sensed (NASA POWER) data for the purpose of estimating monthly total precipitation in Konya Closed Basin. In the modeling process, three different optimization algorithms (ADAM, RMSProp and SGDM) were compared, and it was determined that the ADAM algorithm gave the most successful results in both datasets. In the standard modeling stage performed on the same dataset, using 300 epochs and 10 hidden layers with MGM data, a test RMSE value of 3.52 mm was obtained, with an MAE of 2.59 mm, R² of 0.995 and NSE of 0.980. In the modeling performed with NASA data, the best result was obtained with 300 epochs and 30 hidden layers; test RMSE values of 1.52 mm were reached, with an MAE of 1.06 mm, R² of 0.998 and NSE of 0.995. These results have shown high accuracy and strong model robustness especially in satellite data.

In the cross-validation phase of the study, two scenarios were designed to evaluate both the development and generalization capacity of the model, as well as the independence from data sources. In these scenarios, MGM and NASA POWER were used crosswise: in the first scenario, training was performed with MGM data and testing with NASA data; in the second scenario, the opposite structure was applied. The analyses demonstrated that the model trained with MGM data and tested with NASA data yielded better performance, achieving a test RMSE of 3.62 mm, MAE of 2.93 mm, R² of 0.9966, and NSE of 0.9686 with 300 epochs and 30 hidden layers. In contrast, when trained with NASA data and tested with MGM data, the error rates were slightly higher, with the best results recorded as a test RMSE of 4.48 mm, MAE of 3.22 mm, R² of 0.9921, and NSE of 0.9678 under the same conditions. These findings indicate a high level of agreement between ground-based observations and satellite data, while also highlighting that models trained on MGM data offer better generalization towards remote sensing data. Overall, the cross-validation results confirmed that remote sensing data can effectively replicate ground-based measurements, allowing these two data sources to be used interchangeably under certain conditions. This demonstrates that global datasets like NASA POWER can serve as a reliable alternative in regions with limited data availability or insufficient ground-based observations. While the framework was demonstrated on the Konya Basin, its application to other climatic regions remains a future research priority. Moreover, possible biases in NASA POWER data, particularly in arid zones, should be considered when transferring this approach beyond the study area. Furthermore, this approach supports the development of flexible and sustainable solutions in critical areas such as water resources management and agricultural planning. The detailed comparison of the best-performing algorithm in predicting monthly total precipitation using cross data is presented in Table 5.

The study has shown that LSTM-based deep learning models, especially with the ADAM optimization algorithm, provide a powerful and effective method for precipitation estimation. The ability to use different data sources as alternatives provides significant flexibility in regions where observation infrastructure is insufficient. In addition, it is emphasized that hyperparameters such as the number of epochs and the number of hidden layers should be optimized carefully. The results obtained provide valuable contributions to sustainable water management, agricultural planning and climate adaptation strategies in arid and semi-arid regions. In future studies, it is recommended that this method be applied in different climatic regions and with multivariate datasets, further increasing the estimation accuracy and operational usability. These findings provide practical guidance for sustainable water management and support drought adaptation strategies, particularly in arid and data-scarce regions where reliable precipitation forecasting is critical.