A computer vision approach for satellite-driven wind nowcasting over complex terrains

Accurate wind speed and direction nowcasting in regions with complex terrains remains a challenge, and critical for applications like aviation. This study proposes a new methodology by harnessing Convolutional Neural Networks and Long Short-Term Memory models with satellite imagery to address wind predictions in a complex terrain, centered on Madeira International Airport, Portugal, using satellite data as input. Results demonstrated adeptness in capturing wind transitions, pinpointing shifts up to two hours ahead, with errors of 1.74 m s−1 and 30.98° for wind speed and direction, respectively. Highlighting its aptitude in capturing the intricate atmospheric dynamics of such areas, the study reinforces the viability of computer vision for remote sites where conventional monitoring is either inefficient or expensive. With the widespread availability of satellite imagery and extensive satellite coverage, this method presents a scalable approach for worldwide applications.


Introduction
Accurate forecasting of wind speed and direction is challenging, especially in areas with complex topography such as Madeira Island, where the terrain significantly affects weather patterns, with greater relevance at the Madeira International Airport, Portugal [1,2] which, from this point forward, will be designated by its International Civil Aviation Organization code, LPMA [3].In aviation, precise wind forecasting is crucial for ensuring operational safety, as wind parameters are key factors contributing to weather-related aviation incidents [4].
A suggested solution to enhance forecast accuracy is the increase of numerical weather forecast model resolutions, to account for turbulent wind characteristics and wind shear in the lower troposphere over complex terrains [2].However, these models show limitations, especially in small-scale patterns and physical processes, that are markedly pronounced in complex terrain and pose a risk to aviation [5,6].Machine Learning (ML) techniques, exemplified by models like Convolutional Neural Networks (CNN) that use Deep Learning (DL) networks, emerge as an alternative approach.These DL models learn the patterns directly from the data instead of requiring a complex physical analysis that requires substantial domain knowledge and is hard to adapt to all terrains, thus presenting an alternative to traditional numerical weather models [5,7].Moreover, ML and DL are now surpassing traditional NWP in both result accuracy and operational efficiency.They demonstrate superior processing speeds and reduce computational requirements, further cementing their role as advantageous alternatives in weather forecasting [8,9].
The quality of data used for ML models is a critical determinant of the accuracy of the forecasts generated.High-quality data is vital for training robust models capable of learning the complex dynamics of wind behavior in challenging terrains [10].Recent satellite missions like Aeolus have made significant strides towards better data acquisition by delivering direct observations of wind profile information on a global scale, marking a notable advancement in wind data collection, and providing high-quality data that can be harvested to improve ML approaches [11].
Several studies have explored the use of satellite-derived data for wind speed extrapolation, employing various methodologies.For instance, the use of Atmospheric Motion Vector (AMV) derived from IR10.8 imagery has been demonstrated by Oh et al [12] to be capable of providing accurate measurements.Additionally, the application of Sentinel Application Platform (SNAP) toolboxes for wind speed extrapolation from satellite imagery has been investigated, showing a strong correlation, with a R 2 of 0.87, between wind sensors and satellite-derived wind estimations [13].Wang et al [14] achieved a Mean Square Error (MSE) of 1.22 for wind speed and a 38.74 MSE for wind direction by utilizing satellite backscattered signals.However, the temporal resolution was limited to twice per day, which was the best available at the time.In another study, DL, particularly CNN models, was used for wind speed extrapolation from satellite imagery, achieving an 85% accuracy with 1 h interval measurements, highlighting the applicability in areas such as aviation [15].
Regarding wind nowcasting via satellite data, there are limited studies available.Through a bibliographic search, it was concluded that the existing works primarily utilize SNAP tools to extrapolate wind data, which is then employed as time series for making predictions focused solely on wind speed [16,17].This highlights a research gap where the nowcasting of both wind speed and, especially, wind direction, which is a recognized challenging variable to forecast due to its high variability [18], using satellite imagery and computer vision techniques, has yet to be studied to the best as the authors' knowledge.
It is important to observe that most state-of-the-art works employ data collected locally by wind sensors and then use this data to produce the models, frequently employing ML algorithms that can assess the recurrent patterns, commonly using a Long Short-Term Memory (LSTM) [18], or by producing complex features that can highlight the patterns in the data [1].
Therefore, combining the two approaches presented in state-of-the-art, a comprehensive approach integrating CNN (shown to be capable of examining satellite imagery) and LSTM (proven capable of learning the complex temporal patterns) models with satellite-derived data for both wind speed and direction forecasting in complex terrains has not been thoroughly explored.The use of CNN-LSTM, known for the capability in handling spatial data in image sequences, which along with high-resolution satellite imagery could provide a significant improvement in accurate and reliable wind forecasting in challenging terrains.
The deployment of wind measurement stations in remote and complex terrains presents challenges associated with the high costs and maintenance concerns and most of these sites do not have a wind resource map [19].Therefore, there is an opportunity to the development of systems that can deliver accurate wind data extrapolation and prediction without requiring permanent station or other equipment installations on such complex sites.
This paper proposes a novel methodology, leveraging CNN-LSTM and satellite images for wind speed and direction nowcasting in complex terrains, using LPMA as a case study.The primary objective is to demonstrate that, using computer vision and ML-DL techniques, wind speed and direction can be accurately forecasted solely through satellite imagery, particularly in complex terrains, using real world data, building on Jamaer et al [20] work that proposed the use of cloud information to extract weather data.The global coverage of satellites, along with the wide availability of satellite images, enhances the scalability of the presented solution and its applicability in regions where wind monitoring or forecasting systems may be unavailable deprecated, or inaccurate.
This paper is organized as follows: In section 2, the dataset and methodology employed are described.Section 3 presents the results and provides a discussion.Concluding remarks are extended in section 4.

Data
The data used in this study includes satellite imagery and wind data.The satellite imagery was collected from the '0-degree service', a key mission of Meteosat Second Generation, which furnishes High Rate SEVIRI image data across 12 spectral bands from 01-01-2022 at 00:00 UTC until 31-12-2022 at 23:30 UTC at 30-min intervals, leading to 17,520 images.The SEVIRI instrument facilitates a thorough image scan of the Earth at 15-minute intervals, presenting a high-frequency data acquisition, fundamental in tracking the evolution of cloud structures and patterns in near-real-time, which opens opportunity cloud tracking and, successively, short-term wind speed and direction forecasting [21].
The chosen wavelength data for analysis was the Infrared band centered on 10.8 μm -Channel 9 (IR 10.8).This channel is instrumental in deriving Cloud Motion Vectors (CMVs) [22] by tracking cloud feature movements between consecutive images, which are correlated to wind speed and direction at cloud layer altitudes.Unlike visible bands, the IR 10.8 μm band enables continuous observation both during the day and night, essential for ongoing wind speed forecasting, and is also less impacted by atmospheric scattering and absorption, ensuring stable and consistent measurements beneficial for computer vision techniques [21,22].Figure 1 shows an example of two satellite images from the database with 2 h difference.
The wind data, used as ground truth labels, was acquired from LPMA Meteorological Aerodrome Report (METAR) reports from 01-01-2022 at 00:00 UTC until 31-12-2022 at 23:30 UTC.These reports generated every 30-minutes, include the 10-minute mean values of the observed wind speed (in knots) and direction (in degrees) at the runway, among other weather-related information [23].Data was retrieved from Iowa State University's -Iowa Environmental Mesonet (IEM), directly linked to the National Oceanic and Atmospheric Administration (NOAA) Automated Surface Observing System (ASOS), which keeps a global archive of METAR information.Figure 2 displays LPMA location highlighted by the official wind measuring position (MID), and figure 3 illustrates the wind speed and direction extraction from the METAR information [24].
The final dataset includes 17,504 data points for both wind speed and wind direction, accompanied by a nominal missing data rate of 0.09% for each of these variables.Concurrently, the dataset furnishes 17,511 data points pertaining to satellite images, with a slightly diminished missing data rate of 0.05%.The mean observed wind speed was 4.79 m s −1 with the maximum record getting to 14.91 m s −1 .The prevailing wind direction was from the North (350°to 20°), corresponding to 50.48% of the observed values.

Methods
The initial preprocessing of satellite images involved reducing their resolution from 1920 × 2032 to 224 × 224 pixels, a standard and widely used resolution in Machine Learning [25].This approach not only accelerates the modeling process but also enhances efficiency by minimizing computational demands [26].A subsequent processing step involved the transformation of these images into arrays.The pixel intensity values, originally ranging from 0 to 255, were normalized by dividing by 255.Wind data was extracted from METAR reports.The original wind speed values, recorded in knots, were converted to m/s.Subsequently, these values were decomposed into  u and  v wind vector components, as this transformation has been shown to enhance machine learning predictions for wind speed and direction [1].A comparison was made between the images and labels to ensure congruence, where labels without corresponding images, and images without corresponding wind information were identified and removed from the dataset.
The wind data was normalized using where X represents the wind value, while X min and X max are the minimum and maximum values in the dataset, respectively.This scaler was applied to the  u and  v wind vector components, rescaling their values to the range [0, 1] based on the minimum and maximum values of the data within these columns.Post-normalization, the data was partitioned into training, validation, and test sets.The training set comprised data from January 1, A grid search was conducted to ascertain the optimal kernel size and number of filters for the model.The search commenced with a configuration of one filter and a 2 × 2 pixels window size and concluded with ten filters and a 10 × 10 pixels window size.This strategy was adopted as focusing on optimizing the number of kernels and kernel size in a ConvLSTM2D layer offers computational efficiency and enhances sensitivity to spatial features.This targeted approach efficiently explores key aspects of the model's architecture, allowing for a clearer understanding of its inner workings and promoting simplicity in hyperparameter tuning.In figure 3, an illustration of the model is shown, with [27]: where X t is the input at time step t, - H t 1 is the hidden state from the previous time step, - C t 1 is the cell state from the previous time step, W xy represents the convolutional weights, b b b b , , , and i f o c are the bias terms for the input, forget, output gates, and cell update, respectively, * denotes the convolution operation  , is the Hadamard product (element-wise multiplication), and s and tanh are the sigmoid and hyperbolic tangent functions, respectively.
The convolution, illustrated in the figure 4 as 'C' is a process where each pixel of an image is transformed using a small matrix, called a kernel, to produce a new image, which can be achieved with [28]: where W represents the width and H denotes the height of the input, and k is the dimension of the square kernel, the output's width and height are typically represented by F and Q. Usually, F is equivalent to W, and Q matches H.This assumes that the input has been correctly padded [28].
In the training phase of the model, an early stopping mechanism was implemented to prevent overfitting.This mechanism ceased training when there was no improvement in the validation loss over 10 consecutive epochs.To capture the most effective learning state, the model's weights were reset to those from the epoch with the best performance metric when the training was stopped.
The best model hyperparameters, specifically the number of filters and the kernel size for both wind speed and wind direction, were found by using a method known as grid search.This systematic approach searches over the entire space for hyperparameters, creating all possible combinations and thereby optimizing the exploration of the parameter space.With each parameter treated as having an equal probability of impacting the optimization process, this method effectively identifies the optimal configuration for these specific hyperparameters [29].For this specific work, aimed at predicting both wind speed and wind direction, the strategy deviated from selecting the best result for a single variable.Instead, the results for both variables were overlapped, and the best configuration coinciding in both was chosen.This approach ensured a more tailored and effective model for the dual objectives of the study.
The optimized model was trained using the designated dataset, validated, and subsequently tested.This process involved the input of a sequence of four images, one for each time step, corresponding to a two-hour period, and outputting an extrapolation of the wind at T0, corresponding to the observed time, and four 30minutes steps ahead (T1, T2, T3, and T4), providing a forecast up to 2 h in the future with a granularity of 30minutes.Both wind speed and direction were extrapolated at the mid-point of the LPMA runway.The two-hour period was selected based on the satellite imagery covering a 100 km area around Madeira Island, aligning with the maximum recorded wind speed of 14.92 m s −1 , which equates to approximately two hours between the satellite image border and the LPMA MID measuring point.For testing the model's performance, the Mean Absolute Error (MAE), MSE, and Mean Absolute Percentage Error (MAPE) metrics were used, with consideration to the circular characteristics of the wind direction [1,8], with widely used metrics to assess the performance of ML models in predicting wind speed and direction, with as proposed by Alves et al [18].Figure 4 summarizes the methodology employed, providing a visual representation of each followed stage.

Results and discussion
In figure 5, the analysis of the grid search hyperparameter optimization for the model is presented.The more balanced model was chosen based on overlapping the top 10 results for both wind speed and wind direction and selecting the best configuration that appeared in both variables.This approach led to the selection of a model with 4 filters in a 3 × 3 window.For this hyperparameter configuration, at T0 prediction, the MSE values for wind speed and direction were 8.54 and 4,966.96for the wind speed (figure 5(A)) and direction (figure 5(B)), respectively.This selection process ensures that the chosen model is capable of forecasting both wind speed and direction effectively, without disproportionately optimizing for one variable at the expense of the other, thus achieving a balanced performance across both metrics.Analyzing results shown on figure 6, which reflects the entire dataset wind speed and direction predictions across time intervals T0 through T4, predicted using the tunned model, several noteworthy patterns emerge.With respect to wind speed, both T1 and T3 consistently register superior performance in terms of error metrics, underscoring their enhanced prediction accuracy relative to the other intervals.Conversely, T4 demonstrates higher error values, alluding to its reduced prediction precision.
For wind direction metrics, T1 markedly exhibits the lowest error values across all the metrics, reinforcing its robustness in prediction performance.In contrast, T4 persistently records the highest error values, indicating the higher challenge for predictions accuracy in this interval.
To further analyze the performance, a wind direction analysis was performed by wind quadrant.In figure 7, the MAE variation across each forecasted step for the North, East, South, and West quadrants is illustrated for wind speed (figure 7(A)) and wind direction (figure 7(B)).
The results reveals that the South quadrant demonstrates a relatively higher error rate compared to the others.The MAE in wind speed reaches up to 4.95 m s −1 , and for wind direction, it peaks at approximately 120°.
It is also noticeable that the most equilibrated forecast for speed and direction is for the North quadrant, with a minimum MAE of 1.54 m s −1 and 27.50°for wind speed and direction respectively.A probable reason for this divergence may possibly be the relatively infrequent occurrence of southerly winds in the dataset, constituting only 22.7%.This is consistent with the predominantly northern and northeast winds characteristic of the region [2].Consequently, the limited data on southern wind patterns potentially leads to greater difficulties for the ML model in capturing atmospheric dynamics for this quadrant.
Following the detailed analysis of the entire test dataset (2022-11-01 to 2022-12-31), a one-way analysis of variance (ANOVA) was conducted to statistically evaluate the differences in MAE, MSE, and MAPE across the time intervals T0 to T4.The defined null hypotheses were 'there are no significant differences in the mean errors for MAE, MSE, and MAPE across different time intervals'.The results shown that the null hypothesis can be rejected, with all p-values falling below the 0.05 significance threshold [18,30].This outcome underscores statistically significant disparities in the errors across the intervals, confirming the observed patterns of T1 and T3 exhibiting enhanced performance and T4 revealing a more challenging scenario, not as random variations but as significant trend when considering the entire test dataset.
The observed differences in errors across various forecast intervals, as highlighted by the one-way ANOVA results, do not indicate a simple pattern of increasing error over the forecasted period.As shown in figure 6, the forecast for 1:30 h ahead (T3) exhibit lower errors than the forecast for 30 min ahead (T1).This phenomenon arises from the model's methodology of making independent predictions for each timestep instead of building on previous forecasts.This way, the model utilizes ground truth data to independently predict each timestep.The better convergence for T3 suggests that the sequence of satellite images generally provides more relevant information for the spatiotemporal context of 1:30 h in relation to the LPMA position and indicates that the atmospheric dynamics captured in the images during this specific interval were assimilated effectively by the model.
Figure 8 depicts the wind roses of the true wind speed and direction in comparison to the predictions from T0 to T4, providing insights into the general wind conditions during the entire period from 2022-11-01 at 02:00 UTC to 2022-12-31 at 23:30 UTC.The results indicate that all forecasted periods consistently show the prevailing wind coming from the North, which aligns accurately with the true wind direction.Furthermore, the wind speed predictions are consistent across all forecasted periods.The forecasted mean wind speed of the entire dataset deviates by 1.12 m s −1 from the true value, considering all predicted periods, with a maximum variation of 1.29 m s −1 expressed on T2.This consistency in prediction underscores the reliability of the forecasting model in capturing the wind conditions for the specified period, independently of the prediction period.
Forecasting the wind from the South and Southwest proved to be more challenging.While all forecasted windows were able to predict these occurrences to some extent, T3 displayed the least accurate results in this regard and T1 demonstrated a closer approximation to the true conditions registered, for this quadrant, as can be observed in figure 8.
An essential point to consider is the underlying prediction methodology.The forecasts are principally derived from computer vision techniques analyzing satellite images.The inherent variability and quality of satellite imagery between intervals could influence the prediction efficacy.It is plausible that during the T1 interval, the satellite imagery was more conducive to accurate extraction and prediction of wind patterns, whereas the imagery during T4 was possibly more obfuscated or lacked sufficient patterns for precise forecasting.
To better evaluate the proficiency of the presented methodology, using DL and computer vision techniques in nowcasting wind attributes exclusively from satellite imagery, a granular analysis was conducted over a specified interval and shown in figure 8.This period, spanning from the 17th of November 2022, at 00:00 UTC, to the 19th of November 2022, at 23:00 UTC, is characterized by a marked escalation in wind speed and pronounced variability in wind direction.Within this timeframe, the observed mean wind speed was 4.67 m s −1 , while the wind direction exhibited fluctuations between 0 and 350°.In this sequence, figure 9 displays the true wind speed and direction in comparison to the predictions from T0 to T4.It also provides a detailed view of the satellite images and a differential image for the period with notable wind changes, spanning from 00:00 UTC to 02:00 UTC up to November 18, 2022.The wind is represented by wind barbels, and it is notorious the alignment of the true and predicted conditions, especially at the period previously referenced and marked in the picture by a back square.
In the detailed analysis covering the interval around the 18th of November 2022, a notable shift in wind behavior is evident at midnight.The wind speed accelerates, and the direction stabilizes from the northern sector.Forecasts consistently detect and anticipate these transitions up to two hours prior, aligning accurately with the actual wind recordings.
The satellite images in figure 9 correspond to the referenced period, and it is evident that the cloud movement aligns with the wind direction predicted by the model.The differential image derived from subtracting the pixels of the first image from the last within the timeframe and uses a color scale where red represents 0% difference and blue signifies a 100% difference.This image reveals minimal changes in the southern part of Madeira, which corresponds to the bottom half of the picture.The most significant changes are observed horizontally at the midpoint, indicating that cloud movement was registered in the top half of the image.Therefore, the bluish regions in the differential image from figure 9, indicating areas with more substantial cloud movement, carry the most critical information for analysis by the model, particularly in the top half of the image where these changes are most pronounced.
The model's forecasting accuracy in this period is shown in table 1, where it can be highlighted that the maximum MAE for wind speed and direction was 2.26 m s −1 and 32.36°for the T3 forecasted step, while the minimum was as low as 1.74 ms −1 for the wind speed at the 2 h forecasted step (T4), and 28.83°for the wind direction at T1 step.
These results point out the capacity of the model for correctly forecasting small temporal wind conditions changes.
In an additional case-study, the model's performance in adverse weather conditions, encompassing mist, drizzle, rain, and thunderstorms, was evaluated over a four-day period from 2022-12-01 00:00 UTC to 2022-12-04 23:00 UTC.Meteorological records from the LPMA METAR indicated occurrences of rain and drizzle on the first and third days of this window.The second day featured mist and intermittent rain showers.On the fourth   day, there was significant rainfall and thunderstorms, accompanied by an increase in wind speed from 2 m s −1 to 10 m s −1 .The wind direction rotated from the North-east to the South-east due to the passage of a cold front.During this interval, wind speeds at 300 hPa fluctuated between approximately 10 kt (5.14 m s −1 ) and 100 kt (51.44 m s −1 ), according to the Global Forecast System Analysis records, in the geographic region corresponding to the satellite images used in the model.Table 2 details the model's performance metrics for this specific period.Analysis of these results demonstrates consistency with the overall performance of the model, exhibiting a MAE range from 1.53 m s −1 to 2.05 m s −1 for wind speed predictions, and from 40.30°to 52.09°for wind direction.This underlines the model's ability to deliver accurate forecasts, even in challenging weather conditions and with significant variations in upper atmospheric wind speeds.
As it is evidenced from the general results at T1, from the entire test dataset, which achieved an MAE of 2.19 m s −1 for wind speed and 58.44°for wind direction, the presented approach is robust enough to predict the wind based exclusively on satellite imagery.This effectiveness extends to regions characterized by intricate terrains, such as valleys, mountains, or coastal areas, where wind dynamics are inherently variable, such as Madeira Island [1,2].The model's capability to preemptively identify transitions in wind behavior up to a twohour interval further accentuates its potential utility in wind nowcasting, particularly in remote geographical locations, without the need for permanent local stations or other intricate systems.
Computer vision techniques have been applied in the past for predicting various weather variables [31,32].However, to the best of the author's knowledge, there has been no research adopting this technique specifically for wind forecasting.While machine learning has been employed for wind speed and direction nowcasting using different methodologies, Dupuy et al 2019 [33] stand out as the best-performing work forecasting both variables, by achieving a MAE of 0.54 m s −1 for wind speed and reporting an error within ±45°for 84% of the instances using data from weather stations, as reported recently in a systematic review [18].The effectiveness of these methodologies, although promising, might not align ideally with the approach of the current study.This is due to the complex nature of predicting wind patterns through remote observation of atmospheric or cloud movements, especially when not supplemented by data from locally installed sensors.Nevertheless, in table 3 a summary of relevant studies and their respective results is presented alongside with this work.
Considering the data presented on table 3, other approaches in the field have focused mainly on wind speed prediction.For instance, Oh et al [12], Wang et al [14] and Dupuy et al [33], used various methods like classic algorithms, backscattered techniques, and direct measurements but with longer forecasting steps.James [13], applies SNAP for wind speed prediction with a 2-day forecasting step, and Yayla et al [15], Majidi Nezhad et al [16] and Majidi Nezhad et al [17] employed similar techniques with timeframes ranging from 1 h to several days.While these studies show effectiveness in wind speed prediction, the current study extends the capability to include both wind speed and direction, with a shorter forecasting period, without using locally installed weather equipment, focusing solely on predicting the wind based on satellite information.This novel approach offers a more broad and timely method in the application of satellite imagery for wind forecasting in remote and complex terrains.

Conclusion
Accurate wind speed and direction forecasting, especially in regions with intricate terrains like Madeira Island, is vital for various applications, prominently in aviation for ensuring operational safety.This study presented a pioneering approach, merging the capabilities of CNN-LSTM with satellite-derived imagery, to provide reliable nowcasts of wind speed and direction, using the Madeira International Airport as a reference case.
The major objective of this research was to demonstrate the potential of using computer vision and ML-DL techniques in conjunction with satellite imagery, to predict wind speed and direction, particularly in regions with complex topography, which was successfully achieved, demonstrating capability to nowcast wind speed and direction for the next 2-h, with errors of 1.74 m s −1 and 30.98°respectively.
The novelty of this work lies in its utilization of satellite images as the sole input for wind nowcasting, eliminating the need for permanent local equipment or sensors standing apart from other studies that have traditionally employed weather station data.The integrated methodology of using CNN-LSTM for processing satellite-derived data is innovative, and its efficacy is evidenced by the general results, where the model effectively captured wind behavior changes, preemptively identifying transitions in wind patterns up to 2 h in advance.
The most relevant aspect of this study was the model's ability to predict wind behavior based exclusively on satellite imagery, even in areas with complex terrains, showing its potential to modulate the intricated atmosphere physics of such locations.This introduces significant potential for computer vision application in remote areas or regions where establishing permanent local monitoring stations is infeasible or uneconomical for longer time periods.The scalability of the approach, given the global coverage of satellites and widespread availability of satellite imagery, suggests its adaptability to varied regions worldwide.
The scalability of this model is closely tied to the availability of accurate, localized ground truth data for ongoing retraining and adaptation to diverse terrains and consequent different atmospheric dynamics.As the model expands to new areas, it must be fine-tuned to address varying atmospheric nuances, necessitating a supply of precise, region-specific data.In the absence of this, the model's accuracy could decrease, particularly in regions with distinct meteorological characteristics.Thus, efficient data gathering and processing mechanisms are essential to maintain the model's effectiveness across different geographical landscapes.
Distinct from traditional NWP models, the proposed solution does not require the integration of complex terrain details or other local physical features, nor does it necessitate extensive domain knowledge for its implementation.This aspect significantly simplifies deployment, making it more accessible for widespread use across varied contexts and offering a promising and efficient alternative to conventional forecasting techniques.
This study is limited by its singular focus on one location and the use of one type of satellite imagery.For more comprehensive results, future research should expand to different or even multi locations and explore a variety of satellite images.Additionally, integrating hybrid systems could further enhance the nowcasting accuracy.

Figure 2 .
Figure 2. Madeira international airport official wind measuring position at the midpoint of the runway (MID).

Figure 3 .
Figure 3. Wind speed and direction extraction from METAR [24], showing (A) an overview of the methodology and (B) demonstrating its application at LPMA.

Figure 4 .
Figure 4. Overview of methodological stages and their interconnections.

Figure 9 .
Figure 9. True and predicted wind speed and direction comparison for the period of 2022-11-17 00:00 UTC to 2022-11-19 23:00 UTC and the satellite images with the composite difference matrix for the period from 00:00 UTC to 02:00 UTC of 2022-11-18.

Table 3 .
Relevant studies in the field -comparison with the present work.