Real-Time UAV Scheduling under Hazardous Weather Con-ditions Using an integrated Fuzzy Logic and Machine Learn-ing Framework

The dataset is a hybrid compilation that combines real-time sensor readings with long-term historical meteorological records from Saudi Arabia. Data points were collected at a 30-minute interval, a cadence chosen specifically to capture rapid weather shifts without introducing excessive noise or overhead. Data Sources and Variables (Features)The dataset tracks several core atmospheric factors that directly impact flight safety: Wind Speed: Ranges from 0 to 15 m/s, collected via local ground stations and open-source APIs. Temperature: Ranges from 5 to 45 °C, measured using on-UAV sensors and local weather stations. Rainfall: Ranges from 0 to 12 mm/h, recorded by precipitation gauges and meteorological feeds. Humidity: Ranges from 0 to 100%, obtained from local stations and on-site hygrometers. Flood Probability: A real-time risk index ranging from 0 to 1, integrated from hydrological models. Sandstorm Frequency: Normalized to a 0–1 scale, derived from a 15-year record of sandstorm occurrences and intensities. Engineered Features: Additional features were derived to capture nonlinear interactions, such as wind_x_vis (wind speed multiplied by visibility), pop_x_rain (probability of precipitation multiplied by rainfall), and seasonal flags like is_early_season and is_late_season. Dataset Composition and AugmentationThe total dataset was constructed to ensure the models are exposed to both current conditions and rare, extreme events (like severe sandstorms and floods): Short-term Data: 1,500 real-time samples gathered over a 3-month observation window. Historical Data: 3,500 samples spanning a 15-year historical record. Class Imbalance and SMOTE: Because extreme "Unsafe" weather events are rare (representing only ~5% of the data), the Synthetic Minority Over-sampling Technique (SMOTE) was applied strictly to the training split, augmenting the short-term samples to 2,000. Total Size: After scaling and integration, the total dataset reached approximately 5,500 samples. For the final machine learning training and evaluation phases, 4,500 instances were utilized (3,600 instances for training and 900 for the held-out test set). Labeling and Target ClassesEach sample in the dataset is labeled based on actual flight success logs into one of three feasibility categories: Safe (≈ 80%): UAVs operated normally without any anomalies. Caution (≈ 15%): Flights experienced partial issues such as moderate drift or rapid battery drain. Unsafe (≈ 5%): Flights faced severe risks or resulted in cancellation.