McGuire, MichaelAlmakrami, Khalil2026-06-222026https://hdl.handle.net/20.500.14154/79279Wildfires are becoming more frequent and more severe, which has increased the need for fast detection and reliable short-term spread prediction. In practice, this remains difficult. Satellite systems provide wide coverage, but they can suffer from delay and visibility limits. RGB-based monitoring is faster and cheaper, but it often produces false alarms because smoke and fire can look similar to fog, dust, glare, or steam. In addition, next-day wildfire spread prediction is hard because fire-growth regions are sparse, highly imbalanced, and driven by many geospatial factors. To address these problems, this thesis presents deep learning methods for three connected wildfire tasks: early fire and smoke detection from RGB imagery and next-day wildfire spread prediction from geospatial data. First, ForestFlameNet is introduced as a compact and explainable CNN that uses separable convolutions and attention to support real-time detection. It achieves 98.0% accuracy with only 0.754M parameters. Second, FireSight is proposed to reduce false alarms in difficult outdoor scenes by combining RGB features with Local Binary Pattern and Discrete Cosine Transform cues through gated fusion and imbalance-aware training. On the Wildfire Dataset, it reaches an AUROC of 0.9321 and an F1-score of 82.62%, while keeping deployment cost practical. Third, an Attention U-Net is developed for next-day wildfire spread prediction on the NDWS benchmark using a 12-channel geospatial feature stack. The model improves sparse spread mapping and achieves 0.3134 IoU and 0.4265 Dice, outperforming a standard U-Net baseline. Together, these studies show that efficient design, attention, texture-frequency priors, and imbalance-aware learning can improve wildfire intelligence from early visual alerting to short-horizon spatial forecasting.144en-USWildfire DetectionWildfire PredictionDeep LearningComputer VisionConvolutional Neural Networks (CNN)Attention MechanismsU-NetRGB Image AnalysisGeospatial DataWildfire Spread ForecastingRemote SensingExplainable AI (XAI)Texture and Frequency Analysis (LBPDCT)Thesis