Geostationary Satellite Images Research Articles

Generation of diurnal internal solitary waves (ISW-D) in the Sulu Sea: From geostationary orbit satellites and numerical simulations

Our recent study reported the existence of internal solitary waves with the diurnal tidal cycle (ISW-D) in the Sulu Sea, however, the three-dimensional characteristics and generation mechanism of ISW-D are still unclear (Huang et al., 2023). In this work, the spatial–temporal characteristics, generation mechanism and propagating process of ISW-D in the Sulu Sea are first preliminary investigated based on high-temporal-resolution Geostationary Orbiting Satellite (GOS) images and high-spatial-resolution two-dimensional numerical model (MITgcm). GOS images from 2018-2022 are retrieved and a total of 13 pairs of ISW-D packets are found. ISW-D occur at spring tide during May–August, with an average interpacket distance of 198 km and a phase speed of 2.30 m/s. To further knowledge of the generation mechanism and propagation process of ISW-D, the non-linear and non-hydrostatic numerical simulations are conducted. The comparison of ISW-D parameters with GOS images proves the validity of numerical simulations. The results of numerical simulations and theoretical parameters indicate that ISW-D are generated at the sill near Pearl Bank by internal tide release mechanism. Moreover, three sensitivity experiments are designed to investigate the effects of tidal force and seawater stratification on the generation and propagation of ISW-D. The results reveal that the ISW-D is generated by diurnal tides with stronger intensity than semidiurnal tides. Seawater stratification does not influence the generation of ISW-D but modulates the propagation process. Phase speeds from GOS images and theoretical model show a positive correlation between the phase speeds of ISW-D and the intensity of seawater stratification.

Deep Learning Approach to Improve Spatial Resolution of GOES-17 Wildfire Boundaries Using VIIRS Satellite Data

The rising severity and frequency of wildfires in recent years in the United States have raised numerous concerns regarding the improvement in wildfire emergency response management and decision-making systems, which require operational high temporal and spatial resolution monitoring capabilities. Satellites are one of the tools that can be used for wildfire monitoring. However, none of the currently available satellite systems provide both high temporal and spatial resolution. For example, GOES-17 geostationary satellite fire products have high temporal (1–5 min) but low spatial resolution (≥2 km), and VIIRS polar orbiter satellite fire products have low temporal (~12 h) but high spatial resolution (375 m). This work aims to leverage currently available satellite data sources, such as GOES and VIIRS, along with deep learning (DL) advances to achieve an operational high-resolution, both spatially and temporarily, wildfire monitoring tool. Specifically, this study considers the problem of increasing the spatial resolution of high temporal but low spatial resolution GOES-17 data products using low temporal but high spatial resolution VIIRS data products. The main idea is using an Autoencoder DL model to learn how to map GOES-17 geostationary low spatial resolution satellite images to VIIRS polar orbiter high spatial resolution satellite images. In this context, several loss functions and DL architectures are implemented and tested to predict both the fire area and the corresponding brightness temperature. These models are trained and tested on wildfire sites from 2019 to 2021 in the western U.S. The results indicate that DL models can improve the spatial resolution of GOES-17 images, leading to images that mimic the spatial resolution of VIIRS images. Combined with GOES-17 higher temporal resolution, the DL model can provide high-resolution near-real-time wildfire monitoring capability as well as semi-continuous wildfire progression maps.

Tropical cyclone intensity estimation through convolutional neural network transfer learning using two geostationary satellite datasets

Accurate prediction and monitoring of tropical cyclone (TC) intensity are crucial for saving lives, mitigating damages, and improving disaster response measures. In this study, we used a convolutional neural network (CNN) model to estimate TC intensity in the western North Pacific using Geo-KOMPSAT-2A (GK2A) satellite data. Given that the GK2A data cover only the period since 2019, we applied transfer learning to the model using information learned from previous Communication, Ocean, and Meteorological Satellite (COMS) data, which cover a considerably longer period (2011–2019). Transfer learning is a powerful technique that can improve the performance of a model even if the target task is based on a small amount of data. Experiments with various transfer learning methods using the GK2A and COMS data showed that the frozen–fine-tuning method had the best performance due to the high similarity between the two datasets. The test results for 2021 showed that employing transfer learning led to a 20% reduction in the root mean square error (RMSE) compared to models using only GK2A data. For the operational model, which additionally used TC images and intensities from 6 h earlier, transfer learning reduced the RMSE by 5.5%. These results suggest that transfer learning may represent a new breakthrough in geostationary satellite image–based TC intensity estimation, for which continuous long-term data are not always available.

Accurate nowcasting of cloud cover at solar photovoltaic plants using geostationary satellite images

Accurate nowcasting for cloud fraction is still intractable challenge for stable solar photovoltaic electricity generation. By combining continuous radiance images measured by geostationary satellite and an advanced recurrent neural network, we develop a nowcasting algorithm for predicting cloud fraction at the leading time of 0–4 h at photovoltaic plants. Based on this algorithm, a cyclically updated prediction system is also established and tested at five photovoltaic plants and several stations with cloud fraction observations in China. The results demonstrate that the cloud fraction nowcasting is efficient, high quality and adaptable. Particularly, it shows an excellent forecast performance within the first 2-hour leading time, with an average correlation coefficient close to 0.8 between the predicted clear sky ratio and actual power generation at photovoltaic plants. Our findings highlight the benefits and potential of this technique to improve the competitiveness of solar photovoltaic energy in electricity market.

MT-GN: Multi-Task-Learning-Based Graph Residual Network for Tropical Cyclone Intensity Estimation

A tropical cyclone (TC) is a type of severe weather system that damages human property. Understanding TC mechanics is crucial for disaster management. In this study, we propose a multi-task learning framework named Multi-Task Graph Residual Network (MT-GN) to classify and estimate the intensity of TCs from FY-4A geostationary meteorological satellite images. And we construct a new benchmark dataset collected from the FY-4A satellite for both TC classification and intensity estimation tasks. Four different methodologies to classify TCs and estimate the intensity of TCs are fairly compared in our dataset. We discover that accurate classification and estimation of TCs, which are usually achieved separately, requires co-related knowledge from each process. Thus, we train a convolution feature extractor in a multi-task way. Furthermore, we build a task-dependency embedding module using a Graph Convolution Network (GCN) that further drives our model to reach better performance. Finally, to overcome the influence of the unbalanced distribution of TC category samples, we introduce class-balanced loss to our model. Experimental results on the dataset show that the classification and estimation performance are improved. With an overall root mean square error (RMSE) of 9.50 knots and F1-score of 0.64, our MT-GN model achieves satisfactory performance. The results demonstrate the potential of applying multi-task learning for the study of TCs.

Determination of Low-Intensity Tropical Cyclone Centers in Geostationary Satellite Images Using a Physics-Enhanced Deep-Learning Model

Determination of Low-Intensity Tropical Cyclone Centers in Geostationary Satellite Images Using a Physics-Enhanced Deep-Learning Model

Deriving Severe Hail Likelihood from Satellite Observations and Model Reanalysis Parameters Using a Deep Neural Network

Abstract Geostationary satellite imagers provide historical and near-real-time observations of cloud-top patterns that are commonly associated with severe convection. Environmental conditions favorable for severe weather are thought to be represented well by reanalyses. Predicting exactly where convection and costly storm hazards like hail will occur using models or satellite imagery alone, however, is extremely challenging. The multivariate combination of satellite-observed cloud patterns with reanalysis environmental parameters, linked to Next Generation Weather Radar (NEXRAD) estimated maximum expected size of hail (MESH) using a deep neural network (DNN), enables estimation of potentially severe hail likelihood for any observed storm cell. These estimates are made where satellites observe cold clouds, indicative of convection, located in favorable storm environments. We seek an approach that can be used to estimate climatological hailstorm frequency and risk throughout the historical satellite data record. Statistical distributions of convective parameters from satellite and reanalysis show separation between nonsevere and severe hailstorm classes for predictors that include overshooting cloud-top temperature and area characteristics, vertical wind shear, and convective inhibition. These complex, multivariate predictor relationships are exploited within a DNN to produce a likelihood estimate with a critical success index of 0.511 and Heidke skill score of 0.407, which is exceptional among analogous hail studies. Furthermore, applications of the DNN to case studies demonstrate good qualitative agreement between hail likelihood and MESH. These hail classifications are aggregated across an 11-yr Geostationary Operational Environmental Satellite (GOES) image database from GOES-12/13 to derive a hail frequency and severity climatology, which denotes the central Great Plains, the Midwest, and northwestern Mexico as being the most hail-prone regions within the domain studied.

Using Stereographic Observations of two High Resolution Geostationary Satellite Images for Cloud Detection

Using Stereographic Observations of two High Resolution Geostationary Satellite Images for Cloud Detection

Characterization of Thunderstorm Cells Producing Observable Terrestrial Gamma‐Ray Flashes

AbstractThe meteorological conditions required for the production of Terrestrial Gamma‐ray Flashes (TGFs) are not well understood. Particularly, the link between TGF production, meteorology, and weather severity is poorly characterized with most works focusing on only a small set of TGF events or isolated storms. This work is a further step toward understanding the general context of the meteorological conditions required for TGF production and if it differs from regular lightning production. We use TGFs observed from AGILE, ASIM, Fermi, and RHESSI to generate the largest catalog of TGFs with associated lightning sferics from either the World Wide Lightning Location Network (WWLLN) or Global Lightning Detection (GLD) combined with geostationary satellite images and meteorological conditions derived from ERA5 reanalysis data. In total we analyze 1582 TGF events and contextualize them in comparison to lightning flashes as characterized by ASIM. In our analysis we consider the proportion of TGFs and lightning coming from systems with overshooting tops as well as the Cloud Top Temperature (CTT) and the Convective Available Potential Energy (CAPE). Our results are consistent with previous studies, finding that TGFs observed from space come from primarily higher cloud tops than regular lightning flashes do. We find that CAPE and the proportion of cells with overshooting tops is similar for both TGF and lightning producing cells. It suggests that TGF observations from space are biased toward systems with higher cloud tops because the attenuation of the gamma‐rays from lower altitude TGFs reduce their intensity below the detection level of LEO instruments.

Deep learning methods for intra-day cloudiness prediction using geostationary satellite images in a solar forecasting framework

Accurate solar resource forecasting remains a challenge. Electricity grid applications require both days-ahead and intra-day prediction. Satellite-based methods are known to be the best option for hourly intra-day solar forecasts up to some hours ahead. An adapted Deep Learning (DL) method has been recently reported to outperform the traditional Cloud Motion Vectors (CMV) strategy. This article analyzes the utilization of a well-documented computer vision DL architecture, the U-Net in various forms, for the satellite Earth albedo forecast problem (cloudiness), a straightforward proxy for solar irradiance forecast. It is shown that the U-Net performs better than advanced and optimized CMV techniques and previous art IrradianceNet, setting it at the state-of-the-art. The tests are done over the Pampa Húmeda region of southeast South America, an area in which challenging cloud conditions are frequent. The data for this study are GOES-16 visible channel images. These images present a finer spatial (≃1 km/pixel) and temporal (10 min) resolution than previously explored data sources for solar forecasting. Moreover, the image size used here is ×4 bigger (1024 × 1024 pixels) and the predictions reach further into the future (5 h) than in previous works. The analysis includes several ablation studies, involving different architectures, optimization objectives, inputs, and network sizes. The U-Net is optimized for direct and differential image prediction, being the latter a better-performing option. More notably, the U-Net models are shown to be able to predict cloud extinction, something that has been a barrier for CMV methods.

How useful are the lake surface temperature estimates from a geostationary satellite (Himawari-8) to detect seasonal and diurnal changes?

ABSTRACT The current geostationary satellite images have a time resolution of minutes. It is not clear how useful they are for detecting diurnal and seasonal changes. We tested this by comparing lake surface temperatures (LSTs) from Himawari-8 Advanced Himawari Imager (AHI) images with in situ measurements in Lake Kasumigaura, Japan, for one year. LST products of the Sentinel-3 satellites were also compared. We found that instantaneous LSTs can be estimated with an root mean square difference (RMSE) of 0.6°C. A large amount of data (N = 102–103) is available for each month, but N differs greatly in the range of 216–2771. A good agreement (RMSE of 0.9°C) was found between monthly mean LSTs from Himawari-8 and in situ measurements, and it is better than the RMSE of 1.2–1.7°C found for the Sentinel-3 LST products. The diurnal change detection was more difficult due to insufficient data numbers around noon. Practical remedies for this issue are suggested.

Study on the Slant-Path Effect in the Simulation of Clear-Sky Thermal Radiance for the GK2A AMI

Abstract Taking the slanted satellite viewing geometry into account is important in the simulation of satellite radiances, which vary with the atmospheric conditions along the line of sight. As a first step to take the slanted satellite viewing geometry into account in the numerical weather prediction system operated in the Korean Meteorological Administration, the slant-path modeling is applied for the simulation of clear-sky thermal radiances of a geostationary satellite imager, the Advanced Meteorological Imager (AMI) on board the Geo-KOMPSAT-2A (GK2A). The observations minus simulations (O − B) and the Jacobians before and after the slant-path calculation are compared. Since most infrared channels of AMI are not sensitive to the atmosphere above the tropopause, the size of slant-path effect for AMI is overall smaller than the effect shown in the microwave sounders. Still, the slant-path modeling is found to have a noticeable effect on the three water vapor absorption channels of AMI peaking between 300 and 600 hPa, particularly at large satellite zenith angles and on the regions with high water vapor variabilities in the model field along the line of sight. On average, the slant-path interpolation reduces the standard deviation of O − B of the water vapor channels by around 2.0% on land and 1.4% over the ocean for zenith angles 40°–60°, and it also influences not only the shape and magnitude but also the height of the peak of the Jacobians. In addition, the retrieval experiment based on the optimal estimation also demonstrates the impact of this new approach in the retrieved moisture field, though the improvement is not as significant as shown in the simulation experiment.

Developing a near Real-Time Cloud Cover Retrieval Algorithm Using Geostationary Satellite Observations for Photovoltaic Plants

Clouds can block solar radiation from reaching the surface, so timely and effective cloud cover test and forecasting is critical to the operation and economic efficiency of photovoltaic (PV) plants. Traditional cloud cover algorithms based on meteorological satellite observation require many auxiliary data and computing resources, which are hard to implement or transplant for applications at PV plants. In this study, a portable and fast cloud mask algorithm (FCMA) is developed to provide near real-time (NRT) spatial-temporally matched cloud cover products for PV plants. The geostationary satellite imager data from the Advanced Himawari Imager aboard Himawari-8 and the related operational cloud mask algorithm (OCMA) are employed as benchmarks for comparison and validation. Furthermore, the ground-based manually observed cloud cover data at seven quintessential stations at 08:00 and 14:00 BJT (Beijing Time) in 2017 are employed to verify the accuracy of cloud cover data derived from FCMA and OCMA. The results show a high consistency with the ground-based data, and the average correlation coefficient (R) is close to 0.85. Remarkably, the detection accuracy of FCMA is slightly higher than that of OCMA, demonstrating the feasibility of FCMA for providing NRT cloud cover at PV plants.

Strong Relationship between Eye Radius and Radius of Maximum Wind of Tropical Cyclones

Abstract Estimation of the radius of maximum wind (RMW) of tropical cyclone (TC) is helpful for the disaster prevention and mitigation. If RMWs are estimated from infrared (IR) imagery taken by geostationary meteorological satellites, their estimation is available densely in time, regardless of the ocean basin. Kossin et al. showed that when TCs have clear eyes, the eye radii estimated from IR images have a high correlation with the RMW estimated from aircraft reconnaissance. The regression of the former onto that latter was shown to have a mean absolute error (MAE) of 4.7 km. We revisit the IR-based RMW estimation by using C-band synthetic aperture radar (SAR) sea surface wind estimates. The criteria for selecting clear-eye cases are simplified. The MAE of the Kossin et al. method is found to be smaller than previously suggested: 3.1 km when the proposed relation is used and 2.7 km when the regression is revised with the SAR-measured RMWs. We further propose an improvement of the IR-based method to estimate the eye radii. The resultant MAE is shown to be 1.7 km, which indicates that the IR-based RMW estimation is more accurate than has been suggested. A strong correlation between eyewall slope and eye size is confirmed. We also investigated cloud features in the eye that may be closely related to RMW and wind structure around RMW. Potential applications of highly accurate RMW estimation are discussed. Significance Statement The radius of maximum wind (RMW) of tropical cyclone (TC) is an important factor for TC intensity estimation and disaster prevention. A previous study suggested that the RMWs of TCs with clear eyes can be estimated from geostationary satellite images at a mean absolute error (MAE) of 4.7 km. Here we improved the method, reducing the MAE by more than one-half. Since the method does not require aircraft or satellite in low Earth orbit, it helps TC monitoring at high frequency. The method can also improve initialization of models used to predict TC hazards and further our physical understanding and the climatology of the wind structures near the centers of TCs.

Assessment on the Diurnal Cycle of Cloud Covers of Fengyun-4A Geostationary Satellite Based on the Manual Observation Data in China

Complicated and regionally representative diurnal cycle characteristics of clouds may introduce some errors into the cloud mask (CLM) algorithm of Geostationary (GEO) meteorological satellite imaging system, which are very difficult to be assessed by using analogous products of fixed-passing polar-orbiting satellites. In this investigation, the diurnal cycle of the performance of CLM algorithm of the Advanced Geosynchronous Radiation Imager onboard the China Fengyun-4A satellite (FY-4A/AGRI) is validated by using manually observed cloud covers (CC) at twenty-five ground-based stations in China. The results indicate that the CCs calculated by the FY-4A/AGRI cloud mask algorithm are overestimated at 11:00 BJT (Beijing Time) and 14:00 BJT (around noon) and underestimated at 08:00 BJT and 20:00 BJT (in the morning and evening) at most stations. In summer, compared with other seasons, the cloud covers obtained from the FY-4A/AGRI over northern China and the Tibetan Plateau are much better consistent with the manual observations, but the situation is the opposite in southern China. However, the CC results retrieved at vegetation surface by FY-4A/AGRI show the best and stable performance. Due to that the two independent cloud tests induce most of overestimations, some sensitivity experiments for the CLM algorithm are conducted. The results show that the best improvement effect is achieved after only closing one cloud test using the 3.8 μm band. Many extremely overestimated cloud cover samples (about 56.3%) are eliminated. After that, the FY-4A/AGRI cloud mask product is more reasonable compared with the corresponding infrared and visible imageries.

The retrieval of aerosol optical properties based on a random forest machine learning approach: Exploration of geostationary satellite images

Aerosol optical properties are among the most fundamental parameters in atmospheric environmental studies. Satellite aerosols retrievals that are based on deep learning or machine learning approach have been widely discussed in remote sensing studies, but the flexible random forest (RF) model has not received much attention in the retrieval of geostationary satellite, like Himawari-8. Thus, the Himawari-8 aerosol retrieval achieved by RF model requires further investigation and optimization. Based on the radiative transfer equation, this study proposed a RF model driven by a differential operator, which quantifies a simple linear relationship between aerosol optical depth (AOD) and top-of-atmosphere (TOA) reflectance enhancement. The spectral information of aerosols is achieved by independent TOA reflectance comparison between images rather than one result from multiple band synthesis. The method allows simple feature inputs and shows weak dependence on auxiliary data. It also achieves simultaneous retrievals over different surfaces and maintains mathematical correlation between spectral AODs and Angstrom Exponents (AE). The model performance was evaluated using a series of comprehensive temporal and spatial validation analyses. A sample-based tenfold cross-validation (10-CV) shows that the new method can simultaneously improve the estimation of aerosol properties, with considerably high correlation coefficients (R2) of 0.85 for AODs at the 0.50 μm wavelengths, a mean absolute error (MAE) of 0.08, a root mean square error (RMSE) of 0.13 and >70% of the samples fell within the AOD expected error (EE). The high accuracy of the spectral AOD retrievals also exhibits good performance on AE calculations, with at least 2/3 of the samples falling within the EE. The site based 10-CV also evaluates the spatial predictions on AODs at the 0.50 μm wavelength, with R2 of 0.67, MAE of 0.12 and RMSE of 0.18. It also has outperformed the Himawari operational aerosol products and appeared to be comparable to other popular machine learning models with better AE retrievals in some typical regions. Two typical regional pollution cases also highlight the advantages of the new aerosol monitoring approach. The 5 km resolution aerosol retrievals exhibit good spatial coverage and performance when describing the regional pollution levels and types. The proposed method improves the performance of RF in retrieving aerosol properties from geostationary satellites and also offers a new prospective for aerosol remote sensing using machine learning approaches.

The January 2022 eruption of Hunga Tonga-Hunga Ha'apai volcano reached the mesosphere.

Explosive volcanic eruptions can loft ash, gases, and water into the stratosphere, which affects both human activities and the climate. Using geostationary satellite images of the 15 January 2022 eruption of the Hunga Tonga-Hunga Ha'apai volcano, we find that the volcanic plume produced by this volcano reached an altitude of 57 kilometers at its highest extent. This places the plume in the lower mesosphere and provides observational evidence of a volcanic eruption injecting material through the stratosphere and directly into the mesosphere. We then discuss potential implications of this injection and suggest that the altitude reached by plumes from previous eruptions, such as the eruption of Mount Pinatubo in 1991, may have been underestimated because of a lack of observational data.

FINER-SCALE MONITORING AND ANALYSIS OF 2016–2021 DUST STORMS DISASTERS IN CHINA BASED ON HIMAWARI-8 SATELLITE IMAGERIES

Abstract. Today, remote sensing is one of the important techniques for natural disasters monitoring. Using the Himawari-8 new generation of geostationary meteorological satellite image data, we identified and monitored 2016–2021 dust storm disasters in China at the finer scale, and compared the object-oriented classification, supervised and unsupervised classification methods. Taking the dust storm on May 5, 2016 as example, the average accuracy and kappa coefficient of the supervised classification were 0.78 and 0.68, while the object-oriented classification were 0.81, 0.73. We used the object-oriented classification method to classify the 2016-2021 dust storm incidents across China. In 2021, the largest area affected by sandstorms was about 68.33km2, the smallest area was about 2.08km2, and the average area was about 22.25km2. During 2016–2021, the most severe and wide-spreading sandstorm occurred on March 14–18 2021. The dust storm first appeared in northwestern Mongolia on March 14, and then spread southeast to China. The dust area covered Beijing on the 15th, and then spread from eastern China to the western regions of North China and Northeast China, including most of the Hexi Corridor. On the 16th and 17th, the influence of the dust storms narrowed to central and western Inner Mongolia, northern Gansu and southern Mongolia. The dust storm gradually disappeared on the 18th.

A neural network with spatiotemporal encoding module for tropical cyclone intensity estimation from infrared satellite image

Accurate and instant estimation of tropical cyclone (TC) intensity is crucial for emergency decision making. Although deep neural networks and satellite images have been increasingly adopted for estimating TC intensity, most neural networks that estimate TC intensity from satellite images are largely dependent on the network structure or the quantity and quality of training data, without sufficiently considering the physical characteristics of satellite images and TCs. Therefore, a spatiotemporal encoding module named STE-TC is devised to facilitate estimating TC intensity from the Himawari-8 (H-8) geostationary satellite images, which can enhance spatial features and integrate temporal information of TCs into the satellite images. Furthermore, a TC intensity estimation model named DenseConvMixer is constructed by applying the dense connectivity strategy to an easily stacked ConvMixer network. To demonstrate the effectiveness of the proposed STE-TC module and DenseConvMixer network, a series of experiments and visualization analyses have been conducted. The experimental results show that the STE-TC module can be applied to estimation models based on the convolutional neural network (CNN) to improve their estimation performances. Moreover, the proposed DenseConvMixer model outperforms other CNN-based models on H-8 satellite images, with an average mean absolute error of 6.77 kt and root mean squared error of 8.89 kt.

Solar radiation forecasting with deep learning techniques integrating geostationary satellite images

The prediction of solar radiation allows estimating photovoltaic systems’ power production in advance, guaranteeing a more reliable and stable energy supply. In this work, we present a novel approach for short-term solar radiation forecasting that leverages multi-channel images from the geostationary satellites of the Meteosat series, coupled with GHI values in clear-sky conditions. We propose two distinct deep learning models, a 3D-CNN and a ConvLSTM, to forecast solar radiation in terms of GHI values, up to 6-h ahead with a temporal granularity of 15 min, over a test study area, the city of Turin, Piedmont, Italy. The models have been validated with ground GHI measurements, and the results show that the ConvLSTM consistently outperforms the 3D-CNN for longer forecasting horizons, achieving a MAD of 27.18% and an nRMSE of 0.57 for 6-h ahead predictions. To motivate the use of satellite images, we compared the performance of our approach with a baseline Smart Persistence model and another benchmark model, which previously achieved state-of-the-art performance on the same data set by exploiting various kinds of meteorological inputs. The proposed models outperform the Smart Persistence for predictions farther than 15-min ahead, achieving a Forecast Skill of 0.56 for predictions 6-h ahead. Furthermore, the comparison shows that using raw satellite images overcomes the performance achievable by solely using meteorological variables, reducing the RMSD by more than 3% and the MAD by 1.37% for prediction horizons greater than 4-h ahead.