Temporal and spatial satellite data augmentation for deep learning-based rainfall nowcasting

Precipitation is one of the most difficult components of the hydrological cycle to predict. The forecasting of precipitation is a challenging task due to its spatial and temporal variability, which is influenced by a multitude of factors including pressure, temperature, wind speed, and direction. The accuracy of precipitation forecasting is heavily reliant on the sub-grid scale parameterization of the precipitation formation process. However, the resolution of simulated atmospheric dynamics and sensitivity to initial conditions impose limitations on sub-grid parameterizations, as evidenced by previous studies (Downtown & Bell 1988; Buizza et al. 1999; Harrison et al. 1999; Pappenberger et al. 2005). According to Tian et al. (2019), rainfall forecasting is exceedingly difficult to accomplish in the majority of arid regions.

In addition to the difficulty and intricacy of precipitation forecasting, it is of critical significance in a variety of fields, particularly water sustainability. Natural disasters, such as floods and landslides, are caused by excessive precipitation. The implementation of accurate precipitation predictions enables the adoption of protective and preventive measures against potential disasters, as well as the development of necessary planning (Yildirim & Demir 2022). Insufficient rainfall results in drought, which can lead to scarcities in the food supply, outbreaks of illnesses, and economic setbacks over an extended period (Yesilkoy 2020; Yeşilköy & Şaylan 2022; Islam et al. 2024). To ensure adequate preparedness and strategic foresight, it is imperative to assess both surplus and deficit precipitation levels (Zhao et al. 2018; Gao et al. 2019). The precise prediction of precipitation holds significant importance for the purposes of agricultural irrigation and planning of water distribution (Weesakul & Lowanichchai 2005).

The achievement of adequate food production and the maintenance of its sustainability can only be realized through the methods outlined in previous studies (Dhekale et al. 2018; Tian et al. 2019; Zambrano et al. 2019). The importance of predicting precipitation is widely recognized in the development and arrangement of water distribution systems, as well as in the scheduling of mining operations. The significance of its forecasting in urban areas lies in its profound influence on infrastructure (Alabbad & Demir 2022), traffic, sewerage networks, water infrastructure (Beck et al. 2010), and various other human activities, as highlighted by Hung et al. (2009). Furthermore, the development of flood information systems has gained significant prevalence in recent times (Li & Demir 2022; Yeşilköy et al. 2023). The creation of precipitation prediction models and their integration into these systems has become a crucial undertaking (Baydaroğlu et al. 2023; Sit et al. 2024). Interactive interfaces enable the computation of potential flood scenarios (Sermet & Demir 2022). These scenarios consider river contributions resulting from precipitation, as demonstrated in previous studies (Demir et al. 2015; Zahmatkesh et al. 2019).

Deep learning is an artificial intelligence methodology that has demonstrated significant accomplishments across different fields. The present technique is a machine learning approach that finds application in diverse domains such as biometric authentication techniques like fingerprint and voice recognition, computer vision and natural language processing (Sermet & Demir 2021), iris recognition, and cancer cell identification. The distinguishing characteristic of this approach in contrast to conventional techniques is its superior computational capability, which is attributable to its intricate and multi-layered architecture, as well as its capacity to handle substantial volumes of data. The application of deep learning techniques in the context of environmental issues is a relatively recent development (Bayar et al. 2009; Li & Demir 2023). These methods have been employed in solar energy forecasting by Sun et al. (2019), strawberry production prediction by Chen et al. (2019), ozone estimation by Eslami et al. (2019), estimating water needs by Guo et al. (2018), river water level estimation by Liang et al. (2018), and wind energy forecasting by Wang et al. (2017). Additionally, several researchers including Ayzel et al. (2019), Boonyuen et al. (2018), Akbari Asanjan et al. (2018), Li et al. (2018), Weesakul et al. (2018), and Hernández et al. (2016) have employed deep learning algorithms for precipitation estimation and have achieved successful results.

The convolutional neural network (CNN) is an innovative method that surpasses the constraints of conventional machine learning methodologies in predictive modeling. The formation of feature maps is achieved through the utilization of linear convolution filters in conjunction with nonlinear activation functions within convolution layers. Long short-term memory (LSTM) networks are a type of artificial neural network that were originally introduced by Hochreiter & Schmidhuber (1997) and have since undergone further development through various contributions. There exist memory structures that are conducive to the acquisition and retention of information over extended periods. In the context of artificial neural networks, the ability to establish connections between past information and the current time step is limited when dealing with long-term dependencies. The LSTM model was specifically developed to address this particular issue, as documented by Hochreiter & Schmidhuber (1997) and Olah (2015). Shi et al. (2015) proposed convolutional LSTM (ConvLSTM) as a modified version of LSTM that incorporates the convolution operation within the LSTM cell. This modification aims to enhance the accuracy of spatial data modeling.

Augmenting the quantity of data points, specifically by producing synthetic data, is a valuable strategy for enhancing prediction accuracy and achieving superior analytical outcomes. Data augmentation techniques have become increasingly prevalent (Demiray et al. 2021) in response to the substantial data requirements of machine learning algorithms. The aforementioned techniques encompass conventional spatial (Lucas et al. 2022), temporal (Wang et al. 2020), and spatiotemporal (Chen et al. 2021; Skoulikaris et al. 2022) approaches, in addition to deep learning methodologies (Kumar et al. 2021; Demiray et al. 2023; Sit et al. 2023). The selection of an appropriate data preparation and interpolation technique has critical significance and should be based on the inherent properties of the data (Baydaroğlu et al. 2024). They stated that most interpolation methods produce equivalent outcomes when data are abundant. In mountainous regions, data collection is limited and measurements for certain variables may exhibit significant variations even at spatial scales that are comparatively small (Collins 1995). Several interpolation techniques, including bilinear (Plouffe et al. 2015), bicubic (Peng et al. 2019), nearest neighbor (Lakew & Moges 2021), distance weighted average (Liu et al. 2020), and kriging (Lucas et al. 2022), have been employed to augment precipitation data.

Several studies have been conducted on rainfall prediction using machine learning techniques. In their study, Adaryani et al. (2022) employ a hybrid model, namely PSO-SVR, that integrates particle swarm optimization (PSO) and support vector regression (SVR) techniques, as well as LSTM and CNN approaches, to forecast rainfall depth for both 5- and 15-min ahead. They reveal that the PSO-SVR and LSTM methodologies exhibited comparable or superior performance to that of CNN. Amini et al. (2022) built a deep neural network architecture intended for the purpose of nowcasting rainfall with a lead time of 5 min. To enhance the precision of deep neural networks, they are integrated with the projections of numerical weather forecast models through the application of three ensemble models, namely bagging, random forest, and adaptive boosting. The findings indicate that the ensemble models exhibited a higher level of accuracy, surpassing that of deep neural networks by a minimum of 10% in the majority of rainfall occurrences.

The ConvLSTM-based flood index (Deo et al. 2015) forecasting model was developed by Moishin et al. (2021) to assess the likelihood of flood occurrences at forecast horizons of 1, 3, 7, and 14 days. The findings demonstrate that the flood model based on ConvLSTM outperforms the benchmark techniques. Kumar et al. (2020) introduced a framework for nowcasting short-term precipitation events using satellite data. The model was trained using 10 sets of NASA IMERG precipitation data, each consisting of 30-min intervals. The neural network model that underwent training predicted the 11th instance of precipitation within a series of 10 such instances. Iterative utilization of predicted precipitation data is employed for nowcasting precipitation with a lead time of up to 150 min. The model exhibits accuracy scores of 0.93 and 0.87 for lead times of 30 and 150 min, respectively. Chen et al. (2022a, 2022b, 2022c) devised an LSTM model based on deep learning techniques to predict the distribution of monthly rainfall. The authors conducted a comparative analysis of the LSTM model's performance against a data-driven random forest (RF) model. The findings indicate that the LSTM model exhibits higher accuracy compared to the RF model in predicting the distribution of rainfall.

Chen et al. (2022a, 2022b, 2022c) introduced a two-stream convolutional LSTM architecture, comprising a short-term sub-network and a long-term sub-network, for the purpose of nowcasting precipitation. This model employs collaborative utilization of networks to effectively capture the rainfall data heterogeneity. Moreover, a novel memory cell is devised to capture long-range dependencies spatially and temporally. The model produces quite good outcomes, according to the experimental findings. Chen et al. (2022a, 2022b, 2022c) developed a model for short-term flood prediction. The model incorporates CNN for image processing and LSTM for collecting spatiotemporal characteristics of hydrological data. Interpolated and aggregated rainfall and flow data train the model. They said that the ConvLSTM model exceeded other approaches in three statistical measures.

Rainfall prediction studies that increase the spatial or temporal resolution of the data are limited in the literature. These studies utilized combinations of data acquired from several networks (Huang et al. 2019) or use deep learning techniques (Oza et al. 2020; Harris et al. 2022) for data augmentation. The authors employed eight distinct interpolation techniques to enhance spatial, temporal, and spatiotemporal resolution and assessed the nowcast outcomes.

The intricate and complex characteristics of atmospheric phenomena, which are both nonlinear and multivariable in nature (Baydaroğlu & Koçak 2014), pose significant challenges to the accurate long-term prediction of rainfall. The present investigation involves rainfall nowcast over a period of 12-time steps, which is equivalent to 6 h. It examines two flooding events in the Marmara and Central Black Sea regions in Türkiye through four different datasets and compares the nowcasts' results. The first approach relies solely on raw rainfall satellite-derived rainfall data, while the second, third, and fourth incorporate spatially, temporally, and spatiotemporally augmented data, respectively.

The paper is structured as follows: Section 2 provides an overview of the materials and methods employed in the investigation, Section 3 outlines the results and subsequent discussion, and Section 4 provides a summary conclusion and future suggestions.

This research examines the floods that occurred in the Marmara and Black Sea Regions, utilizing it as case studies. The analysis involves the scrutiny of satellite imagery pre and post the flood event, with a focus on identifying successive images that depict precipitation. NetCDF-formatted data that correspond to satellite imagery captured during the precipitation periods are used to construct the ConvLSTM model. The data, acquired from NASA Giovanni (Goddard Interactive Online Visualization and Analysis Infrastructure), include a multi-satellite precipitation estimate with gauge calibration – Final Run (suggested for general use) (GPM_3IMERgHH_v06) (https://giovanni.gsfc.nasa.gov/giovanni/). The spatial and temporal resolutions of the precipitation data (mm/h) are 0.1° × 0.1° and half hour (30 min), respectively.

The augmentation of data involves the utilization of diverse techniques for spatial and temporal interpolation. The current study utilizes various interpolation methods, including nearest neighbor, bilinear, bicubic, distance weighted average, first- and second-order conservative, and largest-area fraction techniques to augment spatial data. Additionally, linear interpolation is employed to augment temporal data. One of the most straightforward techniques is the nearest-neighbor algorithm, which assigns the value of the nearest existing data point to a new point (Parker et al. 1983). The bilinear interpolation algorithm relies on neighboring points closest to the pixel point, which are then weighted to calculate the average pixel value. The determination of weight is contingent upon the spatial separation between the interpolation point and its closest point (Sa 2014). Bicubic interpolation considers 16 pixels in the closest 4 × 4 neighborhood of known pixels. These are at different distances from the unknown pixel; so, closer pixels are weighted more heavily.

The utilization of bicubic interpolation algorithms results in the production of images that exhibit greater sharpness in comparison to those generated by nearest neighbor and bilinear interpolation techniques (Parsania & Virparia 2016). The distance-weighted average method is a statistical technique used to determine the central tendency of a dataset. In this method, the weighting coefficient for each data point is computed as the inverse of the sum of distances between that data point and all other data points (Burrough et al. 2015). The primary goal of first- and second-order conservative interpolation techniques is to preserve the integral of the data field throughout the interpolation process, spanning from the source to the destination (Shah et al. 2023). The largest area fraction method is predicated upon the identification of the component possessing the greatest proportion of area within each target grid cell (Jones 1998). Linear interpolation involves the computation of a new point that is positioned linearly between two pre-existing points.

Shi et al. (2015) proposed ConvLSTM as a method for precipitation nowcasting. One notable characteristic that sets ConvLSTM apart from other techniques is that its inputs, cell outputs, hidden states, and gates are all represented as 3D tensors, where the final two dimensions correspond to spatial dimensions. The ConvLSTM model predicts the forthcoming state of a specific cell within a grid by utilizing the previous states and inputs of its adjacent neighbors. The aforementioned outcome is attained through the utilization of a convolution operator during the transitions from state-to-state and input-to-state.

The selection of hyperparameters is considered to be the most crucial process in the prediction using deep learning. The optimal hyperparameter varies depending on the specific characteristics of each dataset. Failure to choose the suitable hyperparameter may result in inadequate prediction accuracy or an overfitting problem. The likelihood of encountering overfitting is notably elevated, particularly in predictions derived from limited datasets (Zhang 2018). The problem of overfitting can be defined as an algorithm that tends to learn in a manner that emulates a verbatim replication of the data, thereby deviating from the overarching structure of the data. To address these problems, it is essential to carefully choose appropriate parameters and, in cases where data quality is suboptimal or the quantity of data is insufficient, augment the data points for predictions.

Table 1 presents the hyperparameters utilized in the study. The utilization of the Adam (adaptive moment estimation) optimizer, an algorithm based on gradient descent that integrates the advantageous characteristics of AdaGrad and RMSProp algorithms, has proven to be effective in rainfall nowcasting and the gradual convergence to the minimum of the loss function (Amini et al. 2023). Additionally, it encompasses outliers (Kim & Han 2020). The Nadam algorithm, which combines Adam and Nesterov's accelerated gradient (NAG) algorithms (Timothy 2016; Gad et al. 2021), has been demonstrated to exhibit superior predictive performance compared to Adam in some situations (Halgamuge et al. 2020). The optimizers utilized in the study were Nadam and Adam for Marmara and Central Black Sea regions, respectively.

The Rectified Linear Unit (ReLU) activation function has been observed to mitigate the issue of vanishing gradient. Activation functions such as ReLU and its derivatives, including Leaky ReLU, are frequently employed in the field of deep learning (Kumar et al. 2021; Moishin et al. 2021; Amini et al. 2023). ReLU provides fixed derivatives in positive values that allow the training process to continue even when input values are extreme (Kim & Han 2020). ReLU and Leaky ReLU were used in the model as activation functions. A notable limitation of the ReLU function is its characteristic of having a gradient of zero for inputs that are non-positive. This property renders a considerable number of neurons unresponsive to adaptation, leading to their inactivity and consequent classification as ‘dead’ neurons. This issue can be alleviated through suitable initialization techniques and by employing small learning rates (Zhang 2018). Consequently, learning rates of 0.001 and 0.0001 were utilized.

All batch sizes listed in Table 1 were tested, and the study was subsequently conducted based on the batch sizes that yielded the most optimal outcomes. For instance, in cases where a batch size of 16 is deemed appropriate for processing data, the model may encounter difficulties in its maintenance when the same data are subjected to spatial augmentation, primarily due to insufficient RAM resources. In such cases, reduced epoch numbers, batch sizes, and small learning rates were used to mitigate overfitting issues arising from limited datasets. In addition, a simplistic model was developed to minimize overfitting issues owing to the small datasets.

The 2D map was subjected to convolutions using a filter size of 5 × 5. Three filters were employed, and the resulting output was subsequently processed through the ReLU activation function. Following the ReLU activation function, the resulting output underges additional batch normalization, followed by the application of a dropout rate of 0.1. Subsequently, the following layers executed an identical convolutional operation utilizing filters of dimensions 3 × 3. The mean squared error (MSE) that is commonly employed as a loss function in conjunction with the Adam or Nadam optimizer was utilized. The learning rate was maintained at a value of 0.0001. In addition, all nowcasting tasks were done for a time frame ranging from 1 to 6 h.

In this study, the Python programming language was used with the Keras and TensorFlow libraries. The study utilized Google Colab Pro+ as its computational platform. The ConvLSTM model was trained and validated with Python 3 Google Engine backend (TPU) and 35.24 GB RAM. The study's findings were unable to be obtained in certain instances of utilizing augmented data in spatial and temporal contexts due to insufficient RAM capacity.

One of the primary challenges encountered in the field of rainfall nowcast studies utilizing deep learning techniques is the limited availability of data. This study seeks to investigate the impact of augmenting data using various interpolation methods on the ConvLSTM rainfall nowcast model that has been developed. The present study examines two distinct flood events in the Marmara and Black Sea Regions utilizing satellite data. These regions experienced floods at different times. Furthermore, these geographical areas exhibit distinct orographic formations and are subject to varying air mass influences.

The Marmara Region exhibits the largest population density, which, in turn, leads to significant urbanization. The prolonged and continuous precipitation over a span of 4 days resulted in highly destructive consequences. During the initial stage of the study, satellite data were employed to evaluate the efficacy of the rainfall nowcast model that was developed. Subsequently, the data resolution was enhanced with the implementation of eight interpolation techniques, encompassing spatial, temporal, and spatiotemporal dimensions. The findings were compared by generating rainfall nowcasts using augmented datasets.

Nearest neighbor, linear, bilinear, bicubic, distance weighted average, and largest area fraction, and first- and second-order conservative interpolation techniques were used as data augmentation approaches. The Climate Data Operator tool was utilized in all interpolation calculations (https://code.mpimet.mpg.de/projects/cdo).

The process of augmenting spatial data was conducted by employing various interpolation techniques, including nearest neighbor, bilinear, bicubic, distance-weighted average, largest area fraction, as well as first- and second-order conservative interpolation methods. Figure S1, Supplementary material, presents the rainfall data observed during the time interval from 11:30 to 12:00 a.m. on September 8, 2009. The spatial resolution of the rainfall data provided is 0.1° × 0.1°. In the spatial augmentation component of the investigation, the spatial resolution of the rainfall data was increased via interpolation methods, resulting in a spatial resolution of 0.05° × 0.05°, 0.025° × 0.025°, and 0.0125° × 0.0125° (see Figures A2–A8 S2–S8, Supplementary material). The suitability of bilinear and first-order conservative techniques for interpolating this data is evident from Figures S2–S8, Supplementary material.

In the temporal augmentation component of the study, the temporal resolution of the rainfall data was increased using linear interpolation, resulting in temporal resolutions of 15, 10 and 5-minute compared to the original temporal resolution of 30-minute as shown in Figures S2–S11, Supplementary material.

In the section of the study focused on spatiotemporal augmentation, the spatial resolution of the rainfall data was enhanced by factors 4, 16, and 64 using the aforementioned interpolation methods. Linear interpolation was employed to enhance the temporal resolution of the rainfall data by factors of 2, 3, and 6. Figure S12, Supplementary material, presents the most basic spatiotemporally interpolated dataset, featuring a spatial resolution of 0.05° and a temporal resolution of 15 min.

Some nowcasts that involved the utilization of spatially and temporally interpolated data were unable to be executed due to limitations in the available RAM capacity. Additionally, two nowcasts using the data augmented by bilinear and bicubic interpolation did not yield successful outcomes due to the incompatibility between the model and the data. Table 2 presents the outcomes of the conducted nowcast trials.

Table 2

The outcomes of the conducted nowcast trials for both regions

View Large

The investigation commenced by generating a set of 26 distinct datasets comprising both spatially and temporally interpolated data, as well as raw data. After all nowcast experiments, outcomes were derived from nowcasts generated using a mere seven datasets. The reason contributing to the non-completion of 17 out of the remaining 19 studies was the insufficiency of RAM.

As seen in Figure 5, there are two obvious differences between the nowcast with rainfall data and the nowcast with augmented data: one notable observation is that the training loss and the validation loss exhibit complete convergence in the nowcast utilizing rainfall data. However, this convergence is not observed in the nowcast employing augmented rainfall datasets. Nevertheless, the error indicators suggest that the predictive accuracy of the nowcast generated using augmented data is superior. In the present scenario, when considering the nowcast generated using rainfall data, it may be observed that the model exhibits an overfitting to the training data. Specifically, the ConvLSTM model demonstrates a strong ability to learn from the training data but struggles to effectively generalize its findings to unseen data.

The second is convergence speed. The loss in the nowcast with rainfall data starts with a larger value and decreases rapidly. However, the nowcast utilizing augmented rainfall data exhibits a smaller decrease, although the convergence speed is higher and remains constant beyond the eighth epoch. The interpretation of the ConvLSTM model suggests that it does not exhibit overfitting tendencies toward the training data. In addition, errors of the aforementioned models are comparatively lower than those of the nowcast model incorporating rainfall data. In conclusion, it can be asserted that the utilization of augmented data in nowcast studies leads to enhanced accuracy in predictions. It may be interpreted as meaning that the ConvLSTM model is not overfitting the training data. Moreover, error values are also lower than the nowcast with rainfall data. In conclusion, it can be asserted that the utilization of augmented data in nowcast studies leads to an enhancement in the accuracy of predictions.

Upon careful examination of the nowcast outcomes presented in Figure 7, it is evident that despite minimal rainfall observed at certain points in the northwest of the Marmara Region in the rainfall data (Figure 7(a)), the nowcast results for all methods (both rainfall data and interpolated rainfall datasets) exhibit a broader region with similar rainfall amounts. The key distinction between nowcast results is the tendency of the nowcasting model based on raw data to underestimate rainfall in this region. In other words, the model exhibits a greater degree of predictive accuracy in areas characterized by substantial precipitation as compared to regions with minimal rainfall. This study employed root mean square error (RMSE) and mean absolute error (MAE) as performance criteria, consistent with previous studies (Kumar et al. 2020; Gad et al. 2021; Amini et al. 2022) that utilized these metrics. The error values for the nowcasts are presented in Table 4. According to the findings presented in Table 4, the implementation of data augmentation techniques has been observed to enhance the accuracy of nowcasting. Based on the data presented in the table, it can be observed that the RMSE and MAE values exhibited a reduction of up to 59% when employing augmented data in the nowcast. In all nowcasts generated using augmented data, the error values exhibit a high degree of similarity. There is a lack of substantial disparity in the level of success observed among various interpolation methods.

The Black Sea Region is characterized by being the recipient of the highest amount of rainfall. The Central Black Sea region exhibits the lowest precipitation levels in comparison to its western and eastern counterparts. Notwithstanding this, the duration of the flood spanning 5 days resulted in significant destruction owing to the geographical alignment of the mountains in this area running parallel to the coastline. Additionally, inadequate urban development practices in the beds of the watercourses contributed to the damage. In a manner akin to the research undertaken for the Marmara Region, the present study involved the generation of nowcasts using both raw and augmented data, followed by an analysis of the resultant nowcast outcomes.

Figure S13, Supplementary material, presents the rainfall data observed during the time interval from 10:00 to 10:30 a.m. on August 11, 2021. The spatial resolution of the rainfall data provided is 0.1° × 0.1°. In the spatial augmentation component of the investigation, the spatial resolution of the rainfall data was increased via interpolation methods, resulting in a spatial resolution of 0.05° × 0.05°, 0.025° × 0.025°, and 0.0125° × 0.0125° (see Figures S14–S20, Supplementary material).

In the temporal augmentation component of the investigation, the temporal resolution of the rainfall data was increased using linear interpolation, resulting in temporal resolutions of 15-, 10-, and 5-min compared to the original temporal resolution of 30 min, as shown in Figures S21–S23, Supplementary material.

Linear interpolation increased rainfall data temporal resolution by 2, 3, and 6. Figure S24, Supplementary material, shows the spatiotemporally interpolated dataset, with 0.05° spatial resolution and 15 min temporal resolution.

Table 5 presents the optimal hyperparameter configuration for the Central Black Region flooding, which has been derived from an extensive series of trials.

Figure 10 reveals that while the Western Black Sea Region and central core parts of the Black Sea Region experience no precipitation, the model expects a minimal amount of rainfall in these areas. The estimates of precipitation amounts in other areas of the Black Sea Region are relatively consistent with the statistics on rainfall.

As a result, it can be said that the ConvLSTM model's ability to accurately nowcast flooding in the Black Sea Region is lower when compared to its performance in the Marmara Region. The error values for the nowcasts are presented in Table 6. The analysis of Table 6 indicates that the incorporation of augmented data did not yield any discernible impact on the error values of the nowcast.

The task of predicting rainfall in real time is challenging due to the intricate and ever-changing characteristics of atmospheric phenomena. Numerous research endeavors are currently underway to enhance the accuracy of prediction through the utilization of physical models as well as novel artificial intelligence techniques. However, the escalating frequency of flash floods and severe flooding events, attributed to the effects of climate change, amplifies the significance of this and comparable research endeavors.

This study involved the development of a rainfall nowcast model utilizing a deep learning methodology. The nowcast was subsequently conducted for two distinct flood scenarios, each utilizing distinct datasets which are rainfall data derived from the NASA Giovanni satellite and augmented versions of this data. The objectives of this study were to show how data augmentation impacts rainfall nowcast accuracy and to establish the viability of defining a general rule by conducting a comparative analysis of nowcasts created with the original and supplemented data using deep learning.

The primary challenge encountered during the study pertained to the insufficiency of hardware resources when executing the model on a laptop computer. Despite the utilization of a simple ConvLSTM model and Google Colab Pro + , numerous attempts were made to generate nowcasts using various datasets, necessitating extensive waiting periods for the computation of results.

The Marmara and Black Sea Regions exhibit notable disparities in terms of topography, synoptic patterns, and local orographic characteristics. Due to this factor, precipitation patterns exhibit significant variations across different regions. Hence, these two distinct flood occurrences were selected as case studies to ensure their representativeness.

The quantity of data points plays a crucial role in deep learning methodologies as it enhances the model's ability to effectively learn from the provided data. From this particular standpoint, a total of eight distinct interpolation techniques were employed in order to enhance the resolution of the data in terms of both spatial and temporal dimensions. When examining data augmentation experiments, it is seen that the utilization of second-order conservative interpolation makes interpolating the data the most feasible for the Marmara Region. It is important to acknowledge that while nowcasting with augmented data yielded better results, other interpolation approaches yielded a similar level of enhancement.

Following the completion of the data augmentation phase in the study, a comprehensive evaluation of the hyperparameters listed in Table 1 was conducted. Through this evaluation, the hyperparameters that yielded the most accurate nowcast result were identified. The notable observation at this juncture is that different optimizers yielded superior outcomes, despite employing the identical model and utilizing the same type of data (i.e., satellite-derived precipitation products).

The concluding stage of the study involves generating 6-h nowcasts for 12 temporal intervals, utilizing the hyperparameters that have been determined. The same model was executed using the most suitable hyperparameters for two distinct regions. In the investigation of optimal hyperparameters, it was observed that the rainfall nowcast outcomes exhibited variations across two distinct geographical regions. In the present study, the application of augmented data in nowcast models resulted in a significant reduction of error values in rainfall nowcasts within the Marmara Region. Specifically, the error values were observed to decrease by 59%. Conversely, the utilization of augmented data in nowcast models within the Central Black Sea Region did not yield any discernible impact on the accuracy of the forecasts. Due to the presence of the North Anatolian Mountains, which span across the Black Sea Region in an east-west orientation, orographic precipitation occurs, resulting in substantial levels of precipitation in this area. This study has demonstrated once more that individual rainfall and flood events possess distinct characteristics, including unique structural attributes, synoptic factors, and local features. It is important to note that a model that demonstrates efficacy in accurately predicting a specific rainfall event may not necessarily yield comparable results when applied to another rainfall nowcast. Furthermore, the incorporation of augmented data may not consistently prove useful in every flood event nowcast. In other words, since all flood and rainfall events have different characteristics and are affected by different factors, it may not be feasible to develop a generalized and adaptive rainfall nowcast model.

The application of numerical weather models poses distinct challenges, such as high computational costs and time consumption. Thus, employing machine learning algorithms and, more notably, integrating traditional methods with these algorithms may offer benefits in tackling these challenges.

In order to enhance the scope of this study, it may be advantageous to ascertain the specific type of rainfall, namely orographic, cyclonic (frontal), or convective, and subsequently develop a more intricate model using high-performance computing. To enhance the accuracy of rainfall nowcast studies, it is beneficial to incorporate a precipitation classification stage into the data preparation phase. By identifying the kind of precipitation, the model can be established more effectively, leading to improved prediction outcomes.

The funding for this study was provided by the Scientific and Technological Research Institution of Türkiye (TÜBİTAK), specifically through the 2219 International Postdoctoral Fellowship Program, under Grant Number 1059B191900568.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.