Translating NWP Outputs into UAV-Specific Predictions Using

Machine Learning

David Sládek

Department of Military Geography and Meteorology, Faculty of Military Technologies,

University of Defence, Kounicova 65, Brno, Czech Republic

Keywords: UAV Operations, Numerical Weather Prediction (NWP), Machine Learning, GFS, AROME, ARPEGE,

Drone Weather Forecasting, Local-Scale Prediction, Weather-Aware Autonomy.

Abstract: Unmanned Aerial Vehicles (UAVs) are increasingly deployed in safety-critical, weather-sensitive operations.

However, the direct use of Numerical Weather Prediction (NWP) model outputs often fails to address the

specific operational thresholds and spatial–temporal needs of UAV missions. This study introduces a machine

learning (ML) framework that translates standard NWP forecasts into UAV-specific feasibility assessments.

We integrate both global (GFS) and local high-resolution (ARPEGE, AROME) models to generate real-time,

interpretable indices or GO/NO-GO indicators tailored to UAV performance limits. Our case study over

Nantes (France) for the 2017–2023 period demonstrates the added value of ML-enhanced predictions in terms

of spatial precision, temporal consistency, and decision-support utility. The proposed approach also offers an

effective method to fill gaps in local model availability by learning from global models, ensuring continuity

and operational resilience. By combining observation statistics, NWP forecasts, and ML interpretation, this

methodology supports scalable, automated pre-flight planning under varying weather scenarios.

1 INTRODUCTION

Weather forecasts play a critical role in aviation and

other domains requiring safety-critical, time-sensitive

decisions. In crewed aviation, meteorological

products are issued under strict national and

international guidelines (e.g., WMO, ICAO), with

standardized thresholds, formats, and declared

accuracy metrics. With the increasing use of

unmanned aircraft systems (UAS) for specialized and

autonomous missions, however, the demand for more

precise, localized, and machine-readable weather data

is growing rapidly (Simone et al., 2022).

This shift introduces new challenges. UAS

operations often rely on fine-grained, asset-specific

environmental thresholds—yet conventional

numerical weather prediction (NWP) systems are

inherently coarse in resolution and computationally

expensive. Addressing this gap requires techniques

that can translate general NWP outputs into

personalized, actionable products tailored to a

specific platform or task.

https://orcid.org/0000-0002-4418-3203

Recent advances in machine learning (ML), deep

learning (DL), and IoT offer promising pathways.

These methods have demonstrated success in

enhancing NWP through spatiotemporal pattern

recognition (Ren et al., 2021; Ahmad et al., 2023),

automated statistical post-processing (Rio et al.,

2019), real-time data fusion via IoT networks (Wang

et al., 2022), localized forecasting through

crowdsourced systems (Bindhu, 2020).

While, most studies focus on improving generic

forecast accuracy, operationally relevant

forecasting—e.g., predicting when a UAV can or

cannot fly—remains underexplored. Few works

address how meteorological data can be mapped to

mission-specific thresholds, or how learning models

can be adapted to different assets.

This leads to our central research question:

How well can we predict the personalized operational

limits of specific assets based on various

meteorological data?

This study focuses on learning asset-specific

operational limits from heterogeneous meteorological

184

Sládek, D.

Translating NWP Outputs into UAV-Speciﬁc Predictions Using Machine Learning.

DOI: 10.5220/0013674100003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 1, pages 184-191

ISBN: 978-989-758-770-2; ISSN: 2184-2809

data. Specifically, two integration strategies are

employed: (1) Data-driven ML forecasting, which

learns from historical data or reanalysis (e.g., ERA5)

to generate lightweight, high-resolution predictions

(Ben-Bouallegue et al., 2023), (2) Hybrid ML-NWP

post-processing, where ML enhances assimilation

and post-processing steps (Dong et al., 2023)

Before deploying such systems in aviation, both

benefits and challenges must be considered:

• Benefits: improved accuracy (Sengoz et al.,

2023; Patriarca et al., 2023), faster computation

(Weyn et al., 2020), and enhanced forecast

calibration (Bouallègue et al., 2023)

• Challenges: smoothness bias, difficulty

predicting extremes, integration hurdles, and

regional data inconsistencies (Zhong et al., 2023)

2 DATA AND METHODS

This study builds on the concept of performance

indexes as representations of prediction metadata

(Lombardi et al., 2025) to explore a complete pipeline

for a fully automated weather-based decision support

system. A central challenge is the inhomogeneity of

meteorological data—in format, resolution, content,

and timing.

As Lombardi et al. (2025) note, forecast products

vary significantly across lead times, shifting from

coarse global models to fine-scale local outputs and

finally to real-time observations. We address this with

an ML-based method that supports seamless

transitions between forecast products across planning

and operational phases.

For UAV mission planning, we define seven

categories of meteorological inputs:

• Climate Data – Long-term atmospheric trends

relevant to asset performance.

• Global NWP Forecasts – Medium- to long-

range outputs, up to 21 days ahead.

• Local NWP Forecasts – High-resolution data

available ~72 hours before operations.

• TAF Forecasts – Standard aviation text

forecasts, issued ~24–30 hours in advance.

• Landing Forecasts – Final high-resolution

updates for the approach phase.

• Current Observations – Real-time data from

radar, METARs, or satellites.

• Secondary Products – Crowdsourced or site-

specific data (e.g., webcams, field reports).

This taxonomy enables phase-specific use of

meteorological products, and highlights the need for

adaptive ML models that handle heterogeneous

inputs.

Proposed framework deals with the transition

between the products described in the Table 1 with

basic ML algorithms.

Table 1: Overview of the sources interpreted in this study.

Source Information provided

Climate

Data

Avera

e annual fl

able hours

Optimal asset type selection based on

historical patterns

Global

Model

General operational suitability for a

selected asset (VRFI)

Anticipated energy consumption

(EER)

Local

Model

Forecasts with hi

h spatial resolution

Forecasts with hourl

ranularit

Given the differing nature of Terminal Aerodrome

Forecasts (TAF) and Landing Forecasts provided by

professional meteorologists, as well as nowcasting

methods that rely on real-time observations and

measurements, our current focus is on these available

sources.

2.1 Operational Principles

As UAV operations scale, meteorological systems

must move beyond single-airport support. Our

proposed framework is built on the following

principles:

• Automation – End-to-end, minimal human input;

• Machine Readability – Output in gridded formats

(e.g., NetCDF) for autonomous systems;

• Objectivity – Data-driven decision logic;

• Transparency – Traceable inputs and outputs;

• Relevance – Context-aware, concise outputs;

• Localization – Tailored to local conditions;

• Scalability – Compatible with diverse sources

and missions;

• Human-Like Output – Visuals that support

expert validation;

This framework ensures accurate, scalable, and

interpretable support for automated UAV operations

in an evolving technological and regulatory

landscape.

2.2 Machine Learning Algorithms

We adopt a classification-based ML approach, aiming

to determine whether forecasted conditions are

favourable or unfavourable for a specific asset—a

binary classification problem labelled by finally

Translating NWP Outputs into UAV-Speciﬁc Predictions Using Machine Learning

185

observed conditions. For mapping transitions

between forecast runs and lead times, regression tasks

are used. The selected models span a range of

complexity, interpretability, and robustness, as

summarized in Table 2.

Table 2: Utilised ML algorithms with their expected role in

the study.

Model Remark

Logistic

Regression

Simple, interpretable, effective with

linearly separable data

Decision

Tree

Handles non-linearity;

interpretable; prone to overfitting

Random

Forest

Reduces overfitting, handles

imbalanced data with class weighting

Gradient

Boosting

Powerful, reduces bias, effective

with imbalanced data

AdaBoost

Focuses on misclassified cases,

robust to outliers

KNN

Non-parametric, works well with

small datasets, sensitive to imbalances

2.3 Performance Indicators

Building on previously published work (Lombardi et

al, 2025), we use a framework of indices (Table 3)

that label a prediction based on its relevance to the

flying asset.

Table 3: Selected performance indicators as established by

previous research (Lombardi et al, 2025).

Name Description Value

Vehicle

Feasibility

Index

(VRFI)

Probability that the

predicted values will not

exceed the thresholds

specified for the vehicle

in the mission

0 – surely worse

than threshold

1 – surely better

than threshold

Energy

Efficiency

Rating

(EER)

Estimated energy cost of

flight (direction-

dependent)

0 – min

+∞ – max

These indexes used as labels allow us to assess

both forecast accuracy and operational utility,

bridging physical forecasts with mission-specific

decision-making.

2.4 Observations Statistics

We use METAR reports from Nantes Atlantique

Airport (LFRS) covering 2011–2023, compliant with

ICAO Annex 3. These include wind direction and

speed, visibility, weather phenomena (e.g., fog,

precipitation), temperature and dew point, cloud

cover and ceiling. This dataset serves as:

• Ground truth for model validation

• Basis for computing performance indices (VRFI

and EER)

2.5 NWP Models

We use both global and local NWP models to support

different forecast horizons.

2.5.1 Global Model

The Global Forecast System (GFS) by NOAA (0.25°

× 0.25° resolution) provides up to 16-day forecasts,

ideal for long-range planning. Key It is initialized at

00, 06, 12, and 18 UTC with forecast intervals: 12-

hour steps, then 3-hour steps below 240h horizon.

As noted in the literature (Benjamin et al., 2016;

Baars et al., 2005), effective GFS use requires

attention to: (1) Forecasted values, (2) Temporal

trends and lead-time consistency, (3) Model

reliability and bias characteristics.

Because raw GFS output is complex, we translate

it into probabilistic, threshold-based forecasts,

answering:

“What is the probability this forecast meets

operational limits?”

This aligns with our use of performance indices

(Section 2.3) and supports actionable decision-making.

2.5.2 Local Models

Local models capture mesoscale phenomena critical

to UAV operations. For the Nantes region, we use:

ARPEGE (5 km resolution, Météo-France) and

AROME (1.3 km resolution, convection-permitting)

A key innovation is predicting local model output

from global model input, enabling early

approximation before high-resolution forecasts are

available. This supports:

• Continuous forecasting across model transitions

• Uncertainty quantification before local model

initialization

Though global and local models share similar data

structures (e.g., GRIB, NetCDF), the higher fidelity

of local models improves classification near critical

thresholds.

3 RESULTS

In this section, we present the interpretation of each

data product and demonstrate their integration across

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

186

forecasting horizons, focused on a use case near the

French city of Nantes. The scenario involves

unmanned aerial vehicle (UAV) flight planning under

operational constraints: Wind limits of 10, 12, or

15 m·s⁻¹, Temperature range between −10 °C and

40 °C, Visual Meteorological Conditions (VMC)

required for operation.

These constraints were applied consistently

across observations and forecast models to evaluate

the Vehicle Related Feasibility Index (VRFI) under

both historical and predictive scenarios.

3.1 Observations

The statistical overview offers essential long-term

insight into UAV operational feasibility at the target

site. Instead of presenting separate charts for each

variable (e.g., wind, visibility, temperature), we use

an integrated heatmap approach to visualize overall

operational suitability over time.

The following heatmap (Figure 1) captures

seasonal and diurnal patterns. Darker cells from June

to September indicate fewer weather-related

constraints, while lighter cells in winter—especially

January—highlight adverse conditions in up to 50%

of hours, often due to IMC, strong winds, or freezing

precipitation.

Figure 1: Percentage of hours of flight viability by month

and the three-hour window in Nantes Airport.

Such patterns simulate expert forecaster knowledge:

• Morning fog in colder months from radiation or

advection processes;

• Afternoon convection in summer causing

temporary disruption.

These observations support operational planning

(e.g., identifying optimal months or times) or risk

mitigation (e.g., reserving backup windows or

alternate sites).

Overall, the heatmap serves as:

• A compact, user-friendly climate overview

tailored to UAV limits;

• A static data layer for clustering or meta-model

integration.

At Nantes airport, for example, long-term

unsuitability in January contrasts with generally

favourable summer months. These insights enable

statistical prediction and planning aligned with UAV-

specific thresholds.

3.2 Global Models

To assess medium-range forecasting potential, we

applied both Random Forest and AdaBoost machine

learning models to data from the GFS global model,

focusing on a forecast horizon of 168 hours (7 days)

to 6 hours prior to the intended UAV operation. For

this analysis, we performed the training on 80/20

train/test split dataset from years 2015-2022.

These algorithms were selected based on their

superior performance during initial validation, and

were used to compute the Vehicle-Related Feasibility

Index (VRFI)—i.e., the probability that the mission

will be feasible given forecasted conditions.

Despite the relatively coarse spatial resolution of

the GFS (0.25° grid), the models accurately captured

synoptic-scale signals, including a cold frontal

passage on 2

November 2023 that led to a notable

decrease in predicted feasibility (

Figure

2).

Figure 2: Probability of wind gust lower than 12 m.s-1

meaning feasibility of mission predicted by Random Forest

and AdaBoost 168 to 6 hours before flight.

Both Random Forest and AdaBoost models

exhibited consistent temporal trends, though Random

Forest tended to underestimate feasibility (i.e., more

pessimistic) and AdaBoost was slightly more

optimistic in its predictions.

Despite the high complexity of the situation, these

outputs demonstrate that even one week in advance,

a well-trained model can provide meaningful early

warning to decision-makers, allowing for adaptive

scheduling or contingency planning.

Translating NWP Outputs into UAV-Speciﬁc Predictions Using Machine Learning

187

3.3 Local Models

In this part of the study, we tested whether global

model forecasts (specifically GFS) can serve as

reliable proxies for high-resolution local model

outputs (ARPEGE, AROME), particularly when

local forecasts are unavailable or delayed. The goal is

to bridge spatial and temporal resolution gaps using

machine learning (ML).

We trained regression models (Random Forest,

Gradient Boosting, Support Vector Regression) on

forecasts from 2017–2019, using GFS data to predict

ARPEGE outputs (Figure 4). These models were then

applied to AROME forecasts (1.3 km resolution)

over ~300 grid points in the Nantes region. (Figure 3).

Figure 3: Comparison of Random Forest (blue), Gradient

Boosting (green), and SVM (yellow) algorithms predicting

ARPEGE wind speed predictions in Nantes area based on

the GFS global predictions. (Highest values: Max error,

middle: MSE, lowest values: MAE).

The models performed well, especially for 10 m wind

speed:

• Mean Absolute Error (MAE) was consistently

<2 m·s⁻¹ within a 24-hour lead time.

• Maximum errors rarely exceeded 5 m·s⁻¹, even

up to a week ahead.

• Day-ahead forecasts showed good agreement,

with errors stabilizing around 4 m·s⁻¹.

These results suggest ML-based smoothing of global

forecasts can approximate local outputs with high

fidelity during synoptic stability. Larger

discrepancies were observed during convective

activity, frontal transitions, or terrain-driven

turbulence—primarily due to limitations in input data

resolution, not ML model design.

Despite this, local models remain essential for:

• Vertical profiling and convection-permitting

outputs;

• Finer representation of terrain and

mesoscale features;

• Improved gradients in wind and temperature

fields.

Combined with ML, local models offer a smoother,

more interpretable depiction of atmospheric

conditions. For example, we used wind-related

predictions to estimate UAV energy consumption,

presenting wind influence as a color-contoured field

instead of traditional wind barbs (Figure 4).

On 6 June 2017, this approach captured the

evolution of operational conditions as a frontal

system passed over Nantes. Early in the day, forecasts

indicated generally favorable flying conditions

(Figure 6 left), but by late afternoon, feasible flight

zones were restricted to the eastern urban area (Figure

6 right).

Figure 4: Energy Efficiency Rating for flight to the

northeast in Nantes on 6th June 2017 showing transition of

the headwind in NE part and tailwind in the SW.

However, as the front advanced and passed over

the region, the area with a high probability of safe

UAV operations narrowed—eventually being limited

to the eastern urban area of Nantes (Figure 7).

Figure 5: Vehicle Related Feasibility Index (VRFI) over

Nantes based on ARPEGE model on 06

June 2017 at 00

UTC and 15 UTC. Roads shown in black.

High-resolution AROME forecasts provided even

more detailed insights with hourly resolution. ML

models maintained predictive stability above 97%,

enabling confident classification of GO/NO-GO

decisions based on platform-specific wind thresholds.

For instance, Drone 1 and 3 had a max limit of

10 m·s⁻¹; Drone 4 was configured for 15 m·s⁻¹.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

188

Figure 6: Vehicle Related Feasibility Index (VRFI) for

drone of operational maxima of wind speed 10 m·s⁻¹ (left)

and 15 m·s⁻¹ (right) for 12 and 16-19 UTC during passage

of the frontal system over the area (City district contours

represented by black lines).

These results show that ML-enhanced local

models can effectively track dynamic atmospheric

features and translate them into operationally relevant

outputs. Minor timing offsets between ARPEGE and

AROME were observed, but spatial patterns

remained consistent. The ML confidence scores also

served as a proxy for model reliability, offering

valuable metadata for decision support in UAV

operations.

4 DISCUSSION

As UAV deployments expand across critical

infrastructure, emergency response, and logistics, the

need for predictive tools supporting real-time

operational decision-making continues to grow.

Traditional aviation forecasting relies heavily on

human interpretation of multi-source model data and

TAFs. However, this study contributes to the ongoing

shift toward automated, user-specific forecast

interpretation using machine learning (ML),

particularly via the Vehicle Related Feasibility Index

(VRFI) and Energy Efficiency Rating (EER)

(Lombardi et al., 2025).

Our primary objective was to develop and

validate a modular ML framework that translates

NWP outputs into actionable UAV mission guidance.

Special attention was given to handling forecast

transitions—from GFS to ARPEGE to AROME—

which vary in spatial resolution and update

frequency, posing challenges to consistency. We

addressed this via a predictive chain that classifies or

regresses meteorological data into mission-relevant

outputs and harmonizes across forecast products

using learned statistical relationships.

A secondary goal was to assess whether global

model outputs (GFS) could approximate local model

behaviour (ARPEGE, AROME), providing

continuity during periods when high-resolution data

are unavailable.

Building on the concept of prediction metadata

indices (Lombardi et al., 2025), our method shifts

from raw meteorological values to decision-oriented

scores. Prior works have focused on MOS and short-

term nowcasting (e.g., Baars et al., 2005; Benjamin et

al., 2016), often limited to single-model input or short

lead times. We extend this by integrating multi-tiered

ML pipelines, combining long-range global data with

high-resolution forecasts, while preserving

interpretability through standardized indices.

Our results demonstrate that ML can:

• Predict AROME wind speed from GFS inputs

with good accuracy (MAE < 2 m·s⁻¹),

• Translate raw forecasts into binary GO/NO-GO

decisions, aligned with UAV safety thresholds,

• Detect transitions and instabilities, such as

frontal passages, up to 7+ days in advance using

GFS alone.

VRFI-based predictions showed consistency

across models (Random Forest, AdaBoost), with

model confidence often exceeding 97%, indicating

robustness under synoptic-scale predictability.

Additionally, AROME-based urban-scale forecasts

enabled high-resolution spatial visualizations,

including GO/NO-GO maps and energy efficiency

gradients, supporting user-facing decision tools.

These outcomes support a paradigm shift in

aviation meteorology—from static, generalized

products to dynamic, asset-specific prediction

workflows. The proposed framework enables:

• Real-time translation of forecasts into UAV-

relevant indices,

• Seamless model transitions across forecast

horizons,

• Continuous situational awareness in both

strategic and tactical windows.

Translating NWP Outputs into UAV-Speciﬁc Predictions Using Machine Learning

189

Such capabilities can be embedded into

operational platforms for UAV operators, civil

protection agencies, or airport authorities.

Despite encouraging results, several limitations

remain:

• Convective and mesoscale phenomena led to

greater errors (e.g., peak wind deviations > 4

m·s⁻¹), likely due to coarse input resolution

rather than model limitations.

• The geographic scope was limited to the Nantes

region; generalizability to other climates or

terrains (e.g., mountainous or tropical) requires

further validation.

• While index-level interpretability was achieved,

internal model explainability (e.g., SHAP values,

feature importance) was not explored in depth.

Additionally, while our focus was on forecast-based

decision support, nowcasting remains critical,

particularly in the final 0–2 hours before take-off or

landing. Previous work (Lombardi et al., 2025)

identified this as the most tactically significant

period. Future work will enhance this window using

high-resolution satellite, radar, and in-situ data.

Planned extensions of this research include:

• Transitioning to probabilistic classification,

incorporating additional indices (e.g., Vehicle

Source Reliability Index, VRSRI).

• Integration of ensemble NWP systems (e.g.,

ECMWF-EPS) for enhanced uncertainty

modelling.

• Utilization of the latest geostationary sounding

satellite data for improved verification.

• Applying meta-learning techniques to

dynamically adapt model selection and feature

prioritization by region and mission profile.

5 CONCLUSIONS

This study demonstrated the feasibility of a fully

automated, ML-based framework for forecasting

UAV-operational weather conditions across all

planning phases. By integrating global (GFS) and

local (ARPEGE, AROME) NWP models with

METAR observations, we addressed key challenges

in aviation meteorology, notably the transition across

forecast products with differing resolutions and

update cycles.

Our results show that ML models can reliably

translate raw forecasts into actionable, asset-specific

indices such as the Vehicle Related Feasibility Index

(VRFI) and Energy Efficiency Rating (EER). This

supports consistent decision-making from long-range

planning to short-term execution, even during data-

sparse periods.

Moreover, user-centered outputs like GO/NO-GO

maps and climatological heatmaps enhance the

interpretability and relevance of forecast products for

UAV mission planning.

Despite strong performance in synoptic regimes,

limitations remain in capturing convective and fast-

evolving weather patterns—highlighting the need for

future integration of nowcasting methods and real-

time observational data.

Overall, the proposed framework offers a scalable

foundation for operational, data-driven UAV

forecasting and sets the stage for further research into

adaptive, real-time meteorological decision support

systems.

ACKNOWLEDGEMENTS

The work was supported by the Project for the

Development of the Organization, DZRO VAROPS

Military Autonomous and Robotic Assets of the

Ministry of Defence, Czech Republic.

REFERENCES

Ahmad, F., Tarik, M., Ahmad, M., & Ansari, M. (2023).

Weather Forecasting Using Deep Learning Algorithms.

2023 International Conference on Recent Advances in

Electrical, Electronics & Digital Healthcare

Technologies (REEDCON), 498-502. https://doi.org/

10.1109/REEDCON57544.2023.10150439.

Baars, Jeffrey A. a Mass, Clifford F. (2005). Performance

of National Weather Service Forecasts Compared to

Operational, Consensus, and Weighted Model Output

Statistics. Online. Weather and Forecasting. 2005, Vol.

20, no. 6, pp. 1034-1047. ISSN 1520-0434.

https://doi.org/10.1175/WAF896.1.

Ben-Bouallegue, Z., Clare, M., Magnusson, L., Gascón, E.,

Maier‐Gerber, M., Janoušek, M., Rodwell, M., Pinault,

F., Dramsch, J., Lang, S., Raoult, B., Rabier, F.,

Chevallier, M., Sandu, I., Dueben, P., Chantry, M., &

Pappenberger, F. (2023). The rise of data-driven

weather forecasting: A first statistical assessment of

machine learning-based weather forecasts in an

operational-like context. Bulletin of the American

Meteorological Society. https://doi.org/10.1175/bams-

d-23-0162.1.

Benjamin et al. 2016, A North American Hourly

Assimilation and Model Forecast Cycle: The Rapid

Refresh. Monthly Weather Review, 144, 1669-1694.

https://doi.org/10.1175/MWR-D-15-0242.1

Bhagavathi, S., Thavasimuthu, A., Murugesan, A.,

Rajendran, C., A, V., Raja, L.,Thavasimuthu, R. (2021).

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

190

Weather forecasting and prediction using hybrid C5.0

machine learning algorithm. International Journal of

Communication Systems, 34. https://doi.org/10.

1002/dac.4805.

Bindhu, V. (2020). Design and Development of Automatic

Micro Controller based Weather Forecasting Device.

Journal of Electronics and Informatics. 2020.

https://doi.org/10.36548/jei.2020.1.001.

Bouallègue, Z., Cooper, F., Chantry, M., Düben, P.,

Bechtold, P., & Sandu, I. (2023). Statistical Modelling

of 2m Temperature and 10m Wind Speed Forecast

Errors. Monthly Weather Review. https://doi.org/10.

1175/mwr-d-22-0107.1.

Boukabara, S., Krasnopolsky, V., Stewart, J., Maddy, E.,

Shahroudi, N., & Hoffman, R. (2020). Realizing the

Benefits of AI across the Numerical Weather Prediction

Value Chain. Bulletin of the American Meteorological

Society. https://doi.org/10.1175/BAMS-D-18-0324.A.

Dong, R., Leng, H., Zhao, C., Song, J., Zhao, J., & Cao, X.

(2023). A hybrid data assimilation system based on

machine learning. Frontiers in Earth Science 10.

https://doi.org/10.3389/feart.2022.1012165.

Gao, M., Hugenholtz, C.H., Fox, T.A. et al. Weather

constraints on global drone flyability. Scientific Reports

11, 12092 (2021). https://doi.org/10.1038/s41598-021-

91325-w

Ghirardelli, J., & Glahn, B. (2010). The Meteorological

Development Laboratory’s Aviation Weather

Prediction System. Weather and Forecasting, 25, 1027-

1051. https://doi.org/10.1175/2010WAF2222312.1

Joshi, V., Jadhav, P. (2024). A Comprehensive Study on

Weather Predicting Automation Bot Using Automation

Software and Machine Learning. 2024 Asia Pacific

Conference on Innovation in Technology (APCIT), 1-6.

https://doi.org/10.1109/APCIT62007.2024.10673441.

Lombardi, M., Sladek, D., Simone F., Patriarca, R. (2025).

No more flying blind: Leveraging weather forecasting

for clear-cut risk-based decisions. Online.

Transportation Research Interdisciplinary

Perspectives., Vol. 30. ISSN 25901982. https://doi.org/

10.1016/j.trip.2025.101349

NCEP GFS 0.25 Degree Global Forecast Grids Historical

Archive. Research Data Archive at the National Center

for Atmospheric Research. (2015). In NCEP GFS 0.25

Degree Global Forecast Grids Historical Archive.

National Centers for Environmental

Prediction/National Weather Service/NOAA/U.S.

Department of Commerce, Computational and

Information Systems Laboratory

Patriarca, R., Simone, F., & Di Gravio, G. (2023).

Supporting weather forecasting performance

management at aerodromes through anomaly detection

and hierarchical clustering. Expert Systems with

Applications, 213(Part C), 119210. https://doi.org/10.

1016/j.eswa.2022.119210

Ren, X., Li, X., Ren, K., Song, J., Xu, Z., Deng, K., &

Wang, X. (2021). Deep Learning-Based Weather

Prediction: A Survey. Big Data Research, 23, 100178.

https://doi.org/10.1016/j.bdr.2020.100178

Rio, J., Silva, P., Novo, I. (2019). Automatic Forecasts

Algorithms. Sexto Simposio Nacional de Predicción

"Memorial Antonio Mestre". https://doi.org/10.31978/

639-19-010-0.187.

Sengoz, C., Ramanna, S., Kehler, S., Goomer, R., & Pries,

P. (2023). Machine Learning Approaches to Improve

North American Precipitation Forecasts. IEEE Access,

11, 97664-97681. https://doi.org/10.1109/ACCESS.

2023.3309054.

Simone, F., Di Gravio, G., & Patriarca, R. (2022).

Performance-based analysis of aerodrome weather

forecasts. In 2022 New Trends in Civil Aviation (NTCA)

(pp. 27–33). IEEE. https://doi.org/10.23919/

NTCA55899.2022.9934004

Wang, Y., Ni, K., Wang, X., & Zhu, J. (2022). Design of

Automatic Weather Monitoring and Forecasting

System based on Internet of Things and Big Data. 2022

International Conference on Sustainable Computing

and Data Communication Systems (ICSCDS), 979-

982. https://doi.org/10.1109/ICSCDS53736.2022.9761

041

Wang, Y., Kim, S., Lyu, G., Lee, C., Lee, G., Min, K., &

Kafatos, M. (2023). Relative Importance of Radar

Variables for Nowcasting Heavy Rainfall: A Machine

Learning Approach. IEEE Transactions on Geoscience

and Remote Sensing, 61, 1-14. https://doi.

org/10.1109/TGRS.2022.3231125.

Weyn, J., Durran, D., & Caruana, R. (2020). Improving

Data‐Driven Global Weather Prediction Using Deep

Convolutional Neural Networks on a Cubed Sphere.

Journal of Advances in Modeling Earth Systems, 12.

https://doi.org/10.1029/2020MS002109.

Zhong, X., Yao, Y., Xu, L., Wu, Y., & Wang, Z. (2023).

WRF–ML v1.0: a bridge between WRF v4.3 and

machine learning parameterizations and its application

to atmospheric radiative transfer. Geoscientific Model

Development. https://doi.org/10.5194/gmd-16-199-

2023.

Translating NWP Outputs into UAV-Speciﬁc Predictions Using Machine Learning

191