Evaluation and improvement of PERSIANN-CCS precipitation data at multiple temporal scales in the Bouregreg–Chaouia Basin (Morocco) Open Access

Satellite-based precipitation monitoring addresses challenges like the lack of gauging stations and limited access to data at small temporal scales. This study evaluates the performance of the PERSIANN-CCS product by comparing satellite data with rain gauge measurements in the Bouregreg and Chaouia watershed in Morocco. The database integrates hourly data from PERSIANN-CCS and 18 rain gauges. After filling data gaps, statistical parameters such as R, NSE, bias, RMSE, and MAE were used to assess accuracy. Results show that accuracy varies with time scales and station locations. Correlation (R) and NSE are moderate to high, with low bias and improved RMSE and MAE at longer temporal scales. Satellite products perform better during summer months but exhibit lower accuracy and higher bias during the rainy season. Trend analysis of annual precipitation reveals a general decline, although no statistically significant trends were found. Corrected satellite data show better alignment with observed trends, indicating improved accuracy. Various machine learning models were evaluated to improve estimates, and Random Forest showed the best results, increasing correlation and significantly reducing bias, RMSE, and MAE. This study highlights the potential of corrected satellite data for hydrological and climate applications, demonstrating improvements through machine learning.

Precipitation measurement is a crucial element in water resource management. The traditional method of measuring rainfall at gauge stations, although reliable, only provides a point value, which limits its usefulness for estimating precipitation over a wider area. Additionally, the density of the rain gauge network, the accuracy of the recordings, and the interpolation methods employed can influence the quality of the results (Mishra 2013). Satellite remote sensing offers a promising solution to these challenges. It provides comprehensive images of the observed area, thus overcoming the problem of spatial discontinuity inherent in rain gauge networks. Ground-based radars, while offering good spatial and temporal resolution (4 km × 4 km in 1 h), are costly and require regular maintenance. Satellites, on the other hand, offer a credible and operational alternative, providing extensive spatial coverage and almost immediate data availability. However, their spatial and temporal resolutions can sometimes be insufficient for the scales of hydrological variability (Gascon 2016).

Morocco, with its climate characterized by highly irregular spatial and temporal precipitation, and the scarcity of climate stations, is a concrete example where satellite remote sensing is proving particularly useful. Satellite precipitation monitoring offers extended coverage of areas inaccessible by other means and provides real-time information. Moreover, access to satellite data available since 1998 fills the gap in ground-based data that is crucial for hydrological studies (El Orfi et al. 2020). Several comparative studies on a global scale have analyzed the performance of several satellite products developed to estimate precipitation, including: CMORPH, TRMM, CHIRPS, TMPA, PERSIANN-CCS, and IMERG-GPM. However, these methods are susceptible to errors attributable to various factors, including the nature of the precipitation measurement (Villarini et al. 2009), variations in land surface properties (Kummerow 1998; Bytheway & Kummerow 2010; Rafik et al. 2021), and variations in weather patterns, seasons, and altitudes (Ebert et al. 2007; Maggioni et al. 2016).

Aksu & Akgül (2020) evaluated CHIRPS in Turkey and found that estimates were strongly correlated at 10-day and monthly time scales, while the correlation was weaker for daily estimates. Feidas et al. (2009) found a good correlation between Mediterranean rainfall measured at stations and that estimated by the tropical rainfall measurement mission (TRMM) in spring and summer, with less correlation in autumn and winter. Nashwan et al. (2019) tested three products in the Egyptian desert (gauge-corrected GSMaP, IMERG, and CHIRPS) and found that CHIRPS was the best at estimating rainfall amounts when taking into account heavy and moderate rainfall events, while IMERG was suitable only for heavy precipitation. In Algeria, Babaousmail et al. (2019) evaluated two satellite-derived rainfall estimates (CHIRPS and CMORPH) over five climate zones, finding good agreement with ground measurements at monthly and annual scales.

In the Moroccan context, various satellite products have been examined for their usefulness in water management, drought monitoring, and hydrological modeling. These include Milewski et al. (2015), who worked with annual values for the whole of Morocco and found that the TMPA products outperform the real-time products in all environments, and that the latest algorithm version (3B42 V7) offers better performance, particularly in areas of low rainfall and high altitude and Tramblay et al. (2016), who worked on a watershed in northern Morocco, concluding that TRMM-3B42 v7 was closest to ground measurements, well reproducing monthly flows as opposed to daily scales marked by poor performance. Ouatiki et al. (2017) extended this evaluation to the center of the country, noting that this product had shortcomings at the daily scale but performed satisfactorily at the monthly and annual scales. Seif-Ennasr et al. (2016), in the southwest, highlighted the accuracy of CHIRPS, which closely aligns with ground observations.

Moreover, Saouabe et al. (2020) successfully validated the GPM-IMERG product for flood modeling in the High Atlas of Morocco. Across the country, Gadouali & Messouli (2020) evaluated four major products (TRMM3B42V7, ARC2, RFE2.0, and PERSIANN-CDR), noting a common tendency to underestimate precipitation at low and mid-altitudes. Overall, however, these products well reproduced annual precipitation and the seasonal cycle, providing useful information for water resource management and hydraulic planning. El Bouhali et al. (2020) showed that GPM precipitation from TMPA and PERSIANN-CCS precipitation present the best results. El Alaoui El Fèls et al. (2022) evaluated monthly precipitation in the Tensift basin and found fairly good agreement between field-measured data and monthly precipitation estimates from the CHIRPS product, which improved further with the use of a bias correction model.

Despite these studies, none has examined instantaneous time scales, thus neglecting immediate flash flood phenomena. To this end, this paper aims to evaluate the PERSIANN-CCS satellite precipitation product at multiple temporal scales (instantaneous, daily, monthly, seasonal and annual) over 18 years, using 18 rain gauge stations in the Bouregreg and Chaouia watershed, and applies machine learning model to correct and improve the accuracy of the product's raw rainfall estimates.

The area's climate is Mediterranean, with mild, wet winters and hot, dry summers. Overall, it is classified as a semi-arid bioclimatic zone. It tends to be subhumid in Rabat region and freshly subhumid at the higher altitudes of the Central Massif. Toward the southern edges, it tends to be arid. The climate of the study area is influenced by latitude, altitude and its openness to the Atlantic coast. In general, rainfall decreases with latitude, following a gentle gradient, but increases with altitude, creating a cross gradient from south to north and from the coast to the interior of the country. Average annual rainfall varies from 300 mm in the southern edge of the Plateau zone (Settat-Ben Ahmed), to 450 mm/year at Rabat, reaching almost 750 mm/year in the high-altitude zone (Oulmès) (Ezzaouini et al. 2020).

In this study, meteorological precipitation data from 18 rainfall stations across the watershed were collected from two main sources: direct observed ground precipitation data and downloaded satellite data with an hourly step over 18 years from 2003 to 2021.

The satellite data used are those of the PERSIANN-Cloud Classification System (PERSIANN-CCS), developed by the Center for Hydrometeorology and Remote Sensing (CHRS) at the University of California at Irvine (UCI). This system classifies cloud characteristics based on height, spatial extent and texture variability from satellite images. It provides an almost instantaneous estimate of precipitation on a global scale, with a remarkable spatial resolution of 0.04° × 0.04° (i.e. around 4 km × 4 km in 1 h). The exclusive use of infrared (IR) data as input, together with its unique cloud clustering algorithm, sets it apart from other precipitation monitoring systems.

The choice of this algorithm was made thanks to its availability (Open access) and its high spatio-temporal resolution. This product has also a short time lapse between data retrieval and output, which facilitate having the precipitation rate for every 60 min, and also allows to quickly model and forecast floods. Note that its rich spatial resolution is about 39 times denser than the resolution of PERSIANN and PERSIANN-CDR. The PERSIANN-CCS algorithm provides high-resolution precipitation in near-real time thanks to its near-instantaneous input source and unique architecture. PERSIANN-CCS works by segmenting clouds into patches using morphological and hydrological segmentation techniques, then grouping cloud patches into cloud ‘types’ based on coolness, geometry and texture via k-means clustering, and finally forming relationships between cloud upper air temperature and characteristic precipitation for each cloud type (Nguyen et al. 2019).

Data preparation for this study involved downloading 18 years of hourly rainfall, from 2003 to 2021, all across the Bouregrag–Chaouia catchment. These satellite data were acquired via the CHRS data portal https://chrsdata.eng.uci.edu/ (accessed 01.02.2024).

The observed rainfall measurement data were provided by the Bouregreg and Chaouia Hydraulic Basin Agency (ABHBC). These data include rainfall measurements recorded at 5-min instantaneous time intervals over the period 2009–2016, as well as daily measurements over an 18-year period (2003–2021). The 5-min data were recorded at nine rain gauge stations (the first nine stations in Table 1) selected for their high data quality and continuity, which is crucial for such a fine time scale. The daily measurements, covering 18 years, were recorded at 18 rain gauge stations spread across the study area (Figure 1). Table 1 shows the locations of the rain gauge stations used in this study.

This section summarizes the methodology adopted to evaluate the performance of PERSIANN-CCS on different time scales. Satellite and gauged data, covering the period from 2003 to 2021, were subjected to analysis and pre-processing before being compared. This comparison was then based on the calculation of regression metrics, followed by a correction of the satellite data.

After downloading the hourly rainfall database, in the form of images containing information on each hour's rainfall rate for the PERSIANN-CCS algorithm, it was necessary to clean up and fill in the data gaps. To fill the gaps in the PERSIANN-CCS precipitation time series, a linear interpolation approach was implemented. First, the data were sorted chronologically to maintain the temporal integrity of the processing. This cumulative series was then linearly interpolated, allowing for the estimation of missing values by assuming a regular progression between adjacent points. Finally, precipitation values were recalculated as the difference between successive values in the interpolated cumulative series. This method ensures temporal coherence and minimizes distortions in the data, while avoiding overly complex assumptions regarding the spatial or temporal variability of precipitation. PERSIANN-CCS satellite rainfall data were transformed into time series, with 1-h intervals, using Python scripts in the Google Colaboratory environment, to facilitate the process given the considerable volume of data for each hour from the year 2003 to 2021. The next step was to unify the time scales, as the ABHBC provided us with 5-min interval data. In order to harmonize this data with the PERSIANN-CCS hourly satellite data, we used a temporal data aggregation code, which grouped the samples by hour, and then calculated the corresponding aggregations. Similarly, a 24-h cumulation code was used to analyze daily variations specific to a particular time window, with a starting point at 7 am. For monthly data, the calculation of monthly sums was automated for various sources. Finally, the calculation of annual sums provided a longer-term aggregated view, highlighting annual trends. These codes were designed to calculate hourly, daily, monthly and annual rainfall values of satellite and gauged precipitation, ensuring that satellite data were aligned on the same time scales as gauged data at all time steps.

Assessing the accuracy of PERSIANN-CCS satellite precipitation data involves comparing it with ground-based measurements from rain gauges in the Bouregreg and Chaouia watershed. This comparison uses various statistical indices to evaluate performance and identify relationships between the datasets. The statistical indices employed in this study are summarized in Table 2.

• Correlation Coefficient (R): As noted by Tang et al. (2020), R measures the relationship between satellite-estimated and rain gauge-measured precipitation. It ranges from −1 to 1, where values close to 1 indicate strong agreement, while values near −1 denote opposing trends. An R value above 0.95 signifies a strong correlation.

• Root Mean Square Error (RMSE): RMSE quantifies the variance and mean error magnitude between satellite and gauge data. According to Santos (2014), RMSE ranges from 0 (indicating no error) to infinity, with lower values reflecting better agreement between datasets.

• Relative Bias (R_Bias): R_Bias highlights the tendency of satellite estimates to overestimate or underestimate rainfall compared to ground measurements. Negative values indicate overestimation, while positive values suggest underestimation (De Araújo 2006).

• Nash–Sutcliffe Efficiency (NSE): NSE measures the predictive performance of satellite data compared to rain gauges. As described by Liu et al. (2020) and Wei et al. (2018), NSE values range from −∞ to 1, with 1 indicating a perfect match. Higher values represent better performance.

• Mean Absolute Error (MAE): MAE provides an average of absolute differences between satellite estimates and ground measurements. Lower MAE values signify greater accuracy.

Table 2

List of continuous parameters used in this study

Indicators .	Equation .	Possible values .	Optimal value .
The correlation coefficient (R)		−1 to 1	1
Root mean square error (RMSE)		0 to +∞	0
Nash–Sutcliffe model Efficiency coefficient (NSE)		−∞ to 1	1
Relative RBias		0 to +∞	0
Mean absolute error (MAE)		0 to +∞	0

Indicators .	Equation .	Possible values .	Optimal value .
The correlation coefficient (R)		−1 to 1	1
Root mean square error (RMSE)		0 to +∞	0
Nash–Sutcliffe model Efficiency coefficient (NSE)		−∞ to 1	1
Relative RBias		0 to +∞	0
Mean absolute error (MAE)		0 to +∞	0

View Large

The description of the RF algorithm with a set of training data ‘d’ and ‘n’ can be described as follows:

• Use the bagging algorithm to create k random subsets (d₁; d₂; … ; dk) from the training data set d.

• For each training subset dk, a decision tree model is created.

• Combine the k trees h₁ (x₁); h₂ (x₂); … ; dk (xk) in an RF set and determine the final classification results by aggregating the majority votes of the individual trees.

The RF algorithm consists of the construction of multiple independent decision trees, where each tree contributes to error correction (Breiman 2001). These trees are trained on historical data sets that include both precipitation predictions and actual observations (Wang & Zhang 2017). Once the trees have been constructed, they are combined to form an overall model capable of predicting the corrections needed to improve the accuracy of precipitation products (Breiman 2001). This RF regression approach offers several advantages:

• Ability to handle complex data sets: The RF can handle heterogeneous and multidimensional data and combine multiple decision trees to obtain more accurate and stable predictions, which is crucial for correcting precipitation products (Hastie et al. 2009).

• Handling non-linearities: The algorithm can capture complex non-linear relationships between variables, which is important for modeling often non-linear meteorological processes (Biau 2012).

• Adapting to changing patterns: The RF can adapt to changes in weather data over time, enabling accurate and continuous correction of precipitation products (Qi 2012).

• Robustness in the face of missing or noisy data, which is often the case with precipitation data (Geurts et al. 2006).

• High accuracy and stability of predictions, essential for guaranteeing the quality of corrected data (Breiman 2001).

To put these concepts into practice, Python code has been developed to re-evaluate precipitation estimate data by comparing them with measurements. This code uses the Pandas library to manipulate the data, scikit-learn to create an RF Regression model, and seaborn for visualization. Satellite and observed data are loaded from CSV files, then reshaped into 2D arrays. A regression model is then created and trained on these data, enabling precise corrections to be made to improve the quality of precipitation data.

The analysis of climatic trends, particularly precipitation trends, is essential for understanding the potential impacts of climate change on local hydrological regimes.

A positive value of S indicates an increasing trend, while a negative value reflects a decreasing trend.

An S value close to zero suggests the absence of any significant trend in the data.

• Estimating the variance of S: When the series contains equal or repeated values, the variance of S must be adjusted. A low variance indicates high precision in assessing the trend, while a high variance reflects greater uncertainty.

The Z result is compared to a critical threshold based on the p-value, usually at a 95% confidence level, to determine whether the detected trend is statistically significant (Karmeshu 2012).

• P-value: is a statistical measure that indicates the level of confidence associated with the trend detected in a time series. It assesses the probability that the variations observed are due to chance, under the null hypothesis (H0) that there is no monotonic trend in the data.

  

If p < 0.05: The trend is statistically significant, and the null hypothesis H0 is rejected. The conclusion is that there is a monotonic trend (increasing or decreasing) in the data.

If p ≥ 0.05: The trend detected is not statistically significant. This means that there is insufficient evidence to reject H0, and the observed trend may be due to chance.

• b. The Sen slope: is a robust non-parametric method used to estimate the amplitude of a trend detected in a time series. Unlike conventional parametric methods, the Sen slope is insensitive to outliers, making it particularly useful for hydrological data often influenced by extreme events (Sen 1968). This method is based on calculating the slopes between each pair of points in the data series. The steps in the method are as follows:

• Calculation of all possible slopes: For each pair of points (x_i, x_j) in the time series, where i < j, the slope between these two points is calculated according to the formula:

  

This operation is performed for all possible pairs of points in the series (Sen 1968). Positive slopes indicate an increase in precipitation over time, while negative slopes reflect a decrease.

• Median of slopes: Sen's slope is the median of all calculated slopes. By taking the median, we minimize the impact of extreme values and measurement errors, giving this method particular robustness. This median slope represents the estimated mean annual variation in the data, and its sign (positive or negative) indicates the direction of the trend (increase or decrease) (Sen 1968).

The combination of the results of the Mann–Kendall test and Sen's slope not only detects the presence of trends in the data, but also enables the magnitude of these trends to be estimated robustly and reliably, even in the presence of outliers or non-linear variations.

Precipitation data are used at different scales, depending on their purpose. In this study, PERSIANN-CCS data were compared with gauged data at all scales: instantaneous, daily, monthly, seasonal and annual.

Quantitative evaluations of satellite precipitation data compared to gauge station data at instantaneous and daily scales are presented in Table 3 for the instantaneous scale and Table 4 for the daily scale. At the instantaneous scale, results reveal a low correlation (R = 0.07), suggesting difficulties in accurately capturing instantaneous precipitation variations. However, the model demonstrates acceptable accuracy with an RMSE of 0.6, highlighting the importance of instantaneous data in describing point weather events. Despite a high RB (90.55), the relatively low MAE (MAE = 0.073) highlights manageable absolute errors. At the daily scale, performance varied between stations, with moderate correlations, but significant relative differences in some cases. The results reveal challenges in model fitting, with negative NSE at several stations, indicating a sub-optimal fit between gauged and satellite data. Despite this, some stations show a good fit with high correlations and minimal systematic errors. These findings are consistent with those of Hsu et al. (1997), who reported similar difficulties in using PERSIANN-CCS for short-duration precipitation estimation. Similarly, Tan et al. (2021) observed that in North Africa, satellite precipitation products struggle to capture localized rainfall events, leading to high biases and low correlations. This can be attributed to both spatial variability and temporal misalignment between satellite data and ground observations (El Orfi et al. 2020). Additionally, the spatial variability of precipitation and the sparse distribution of rain gauge stations likely contribute to these discrepancies, underscoring the challenges of accurately estimating precipitation at finer temporal scales (Tramblay et al. 2016). The negative NSE values observed at several stations are a direct consequence of these biases, as they reflect the inability of the satellite product to consistently represent localized precipitation dynamics (Ait Dhmane et al. 2023). This highlights the need for advanced correction methods to address these biases and improve alignment between satellite estimates and ground-based observations.

At larger time scales, such as the monthly scale (Table 5), model performance shows some improvement over the instantaneous and daily scales, with higher correlations of up to 0.64, a reduced bias of 5.94, and reduced systematic errors at several stations. However, challenges remain, including negative NSE in some cases, and an increased MAE of up to 23 in some stations, underlining the need for a thorough evaluation for each station. As a result, the accuracy and consistency of satellite product data has improved on seasonal time scales.

At the seasonal scale (Table 6), the results reveal specific trends for each station. Some stations exhibit high correlations, with correlation coefficients reaching R= 0.90, while others show significant systematic errors. Table 7 shows the correlation between the two data sources, but this time on an annual scale, with model performance showing moderate to high correlation R= 0.80. These findings align with those of Nguyen et al. (2018), who demonstrated that PERSIANN-CCS achieves better agreement with ground observations at monthly and seasonal scales. Similarly, Rahman et al. (2022) found that despite its limitations at finer temporal resolutions, PERSIANN-CCS performs reasonably well for long-term precipitation monitoring. In Moroccan arid regions, El Alaoui El Fels et al. (2022) concluded that CHIRPS provides reliable estimates at seasonal and monthly scales, with correlation values reaching 0.8 in some cases.

Despite these improvements, systematic errors persist at certain stations, affecting the accuracy of annual precipitation estimates. While results show a general agreement between gauged and satellite data, discrepancies remain, highlighting the need for further refinements to enhance the precision of satellite-based precipitation estimates.

These findings are consistent with previous studies on satellite precipitation performance in arid and semi-arid regions. Salih et al. (2022) reported improved correlations for PERSIANN-CCS-CDR at larger temporal scales in the Tensift basin. Similarly, Milewski et al. (2015) found better performance of satellite products over extended accumulation periods in dry climates. However, Ouatiki et al. (2017) noted that systematic errors at some stations still affect model accuracy, requiring further refinements.

On the instantaneous scale, where initial accuracy was limited, post-correction results showed a marked improvement. The correlation rose from 0.07 before correction to 0.2 after correction (Figure 4 and Table 8), while the RMSE rose from 0.6 to 0.4 after correction, demonstrating a reduction in errors, and the NSE coefficient rose from −3.06 before correction to 0.014, underlining a slight improvement in performance. RB decreased from 90.55 to 0.72 after correction, confirming a correction of the initial bias and an improved representation at this scale. Similarly, at the daily scale (Figure 5 and Table 9), improvements were remarkable, with stations showing moderate to high correlations after correction, as well as reduced systematic errors. For example, at some stations, correlation increased from R= 0.31 before correction to R= 0.43 after correction, with a reduced RB of 0.17 and a positive NSE of 0.18, underlining a better match between gauged and satellite data after correction. Despite these improvements, residual biases remain, reflecting the inherent challenges of estimating precipitation at fine temporal scales, even after correction. These improvements are consistent with findings from El Alaoui El Fels et al. (2022), who observed a significant reduction in bias after applying correction methods in Moroccan arid regions. Similarly, El Orfi et al. (2020) found that bias correction techniques improve the agreement between satellite and ground-based precipitation measurements in the Oued Oum Er Rbia watershed.

The monthly results (Figure 6 and Table 10) showed a clear improvement, with robust correlations and reduced errors after correction. For example, in some stations, the correlation increased from R= 0.61 before correction to R= 0.79 after correction, with a reduced RB of 0.37 and an NSE of 0.61, highlighting better agreement between gauged and satellite data on a monthly scale. Similar enhancements were reported by Salih et al. (2022), who noted that bias correction significantly improves the reliability of satellite precipitation products for hydrological applications.

At the seasonal scale (Figure 7 and Table 11), the model performance improved further. In some stations, correlation increased from R= 0.94 to R= 0.96, with minimal systematic errors (RB = 0.60) and an NSE of 0.78, confirming better agreement between gauged and satellite data. Moreover, RMSE decreased from 10.69 to 8.45, while the NSE coefficient increased to 0.78, indicating a substantial reduction in initial errors.

A similar trend was observed on an annual scale (Figure 8 and Table 12), where the correlation improved from R= 0.79 to R= 0.91, with minimal systematic errors (RB = −1.17) and a significantly enhanced NSE of 0.79, demonstrating excellent agreement between corrected data and annual scale observations. Additionally, the post-correction RMSE decreased from 48 to 32, and the NSE coefficient rose to 0.85, further confirming the reliability of the bias correction.

Overall, the correction of satellite data using gauge data led to substantial improvements in precipitation estimates, as evidenced by the statistical indices. These findings reinforce the importance of bias correction in satellite precipitation products and align with studies emphasizing the need for region-specific adjustments to enhance accuracy. Tramblay et al. (2016) and Ouatiki et al. (2017) highlighted similar challenges in Morocco, stressing the importance of localized corrections to optimize satellite precipitation estimates in semi-arid regions.

In contrast, raw satellite data show higher slopes (−9.75 to −16.0 mm/year) than observations, with significant trends at stations such as Aguibat (Z= − 2.10, p = 0.036) and Tssalat (Z= − 2.17, p = 0.030). These results illustrate an overestimation of climatic variations by raw satellite data, due to the increased sensitivity of satellite sensors to extreme events such as torrential rains, underlining the need for corrections. Corrected data, on the other hand, show moderate slopes (−1.96 to −10.48 mm/year), often closer to observations, with a reduction in extreme fluctuations. For example, at the Ain Loudah station, the corrected slope is −7.40 mm/year (Z= − 1.01, p = 0.31), compared with a raw slope of −8.50 mm/year (p = 0.046), while the Feddan station shows a corrected slope of −10.48 mm/year (Z= − 1.74, p = 0.08).

The corrections thus reduce the bias of the raw data and bring the trends more into line with those of the observed data, although they are not yet perfect. The graphs show a positive correlation between the three types of data, where increases or decreases in observed precipitation are followed by similar variations in the raw and corrected satellite data. However, the differences are smaller for the corrected data, indicating a better match. These results support the objectives of the study by validating the reliability of the PERSIANN-CCS corrected data, providing critical information on the temporal evolution of precipitation and highlighting the need for improvements, notably through the integration of machine learning models to reduce residual biases and refine trends. The analysis of precipitation trends in the Bouregreg–Chaouia watershed highlights a generalized decrease in precipitation, albeit insignificant in the majority of cases, as well as a notable improvement in corrected data, which better aligns trends with observations and offers increased utility for future hydrological and climatic applications.

The results of this study provide strong evidence for the effectiveness of the RF correction in enhancing the accuracy of PERSIANN-CCS precipitation estimates. The improvements observed across all time scales highlight the potential of bias correction techniques in refining satellite-based precipitation data for hydrological applications.

A comparative analysis of adjusted versus unadjusted data revealed substantial biases in the raw PERSIANN-CCS estimates, particularly at fine temporal scales (instantaneous and daily). High RMSE and MAE values indicated significant discrepancies from observed precipitation, limiting the direct usability of uncorrected data in hydrological modeling.

After applying the RF correction, a notable reduction in biases was observed, with lower RMSE and MAE values, indicating improved precision and reduced forecast errors. The corrected data also exhibited higher correlation coefficients and improved NSE, demonstrating a better fit with observed measurements. These improvements emphasize the importance of bias correction techniques for enhancing the performance of satellite-based precipitation estimates in hydrological modeling.

Beyond the statistical improvements, these findings contribute to the ongoing scientific discussion on the use of satellite precipitation products, particularly in regions with sparse ground-based measurements. Bias correction enhances the applicability of satellite-derived estimates, bridging gaps in precipitation monitoring and supporting decision-making in water resource management.

In the context of climate change, where extreme weather events and hydrological variability are increasing, accurate precipitation data is crucial for predicting and mitigating flood risks. This study reinforces the need for improved satellite precipitation estimates, particularly in semi-arid and Mediterranean regions. By demonstrating the potential of correction techniques to refine satellite data, this research supports the development of more reliable hydrological models that can better capture precipitation dynamics under changing climatic conditions.

However, some limitations remain. The reliance on a single satellite product, PERSIANN-CCS, restricts the scope of validation. A multi-source approach incorporating datasets such as IMERG, CHIRPS, and TRMM could enhance accuracy and robustness. Additionally, applying this methodology to other regions or watersheds with different climatic and hydrological conditions would allow for a broader assessment of its effectiveness and generalizability. Future research should also explore advanced correction methods, including deep learning, to further reduce biases and improve data reliability. A comparative analysis with other bias correction techniques and larger datasets would provide deeper insights into the strengths and limitations of different approaches.

This study evaluated the PERSIANN-CCS satellite precipitation data against measurements from 18 rain gauge stations in the Bouregreg and Chaouia watershed, Morocco, across various temporal scales, including instantaneous, daily, monthly, seasonal, and annual scales. The results demonstrated that satellite data aligned more closely with ground observations at coarser temporal scales (e.g., monthly, seasonal, and annual), while their accuracy decreased at finer scales (instantaneous and daily).

Satellite products like PERSIANN-CCS showed strong potential in capturing extreme precipitation values and spatial patterns, making them valuable for water resource management applications. However, their precision remains insufficient for high-resolution applications, such as real-time flood management. The application of the RF algorithm for bias correction significantly improved the quality of satellite data, particularly at coarser temporal scales. Nevertheless, the corrected data still fell short of achieving satisfactory accuracy at instantaneous scales, highlighting a major limitation for real-time flood modeling.

Regarding precipitation trend analysis, observed data indicated a moderate but statistically insignificant decline, while raw satellite data exhibited more pronounced and statistically significant trends, likely due to the increased sensitivity of sensors. After correction of the estimates provided by the PERSIANN-CCS product, its trends became more consistent with ground observations, although residual biases persist, which is to be expected. This highlights the utility of corrected satellite data in capturing local climatic dynamics while emphasizing the need for continuous improvements in satellite products. Future studies should explore additional satellite products and integrate diverse data sources to reduce input biases. Such efforts are critical to enhancing the precision and reliability of satellite-based precipitation estimates for hydrological and climate-related applications.

We extend our heartfelt gratitude to the Bouregreg and Chaouia Hydraulic Basin Agency for generously providing the essential dataset that was crucial for the completion of this study. Additionally, we are deeply grateful to our research team members and our supervisor for their invaluable support and guidance throughout this project. We also thank the editor and reviewers for their time and effort.

This work was supported by the Moroccan Scientific Research-National Center for Scientific and Technological Research (CNRST) under the Research Excellence Grants Program [1 EHTP2022].

All relevant data are available from an online repository or repositories: https://chrsdata.eng.uci.edu/.

The authors declare there is no conflict.