The growing incidence of agricultural fires in arid and semi-arid regions necessitates the development of advanced, scalable systems for early detection. This study presents FlameGuard, a geospatial AI framework developed to classify fire confidence levels using thermal satellite imagery-based data from the FIRMS dataset and multi-model predictive architectures. The framework integrates a spectral rule-based model—Spectral Algorithm for Fire Detection (SAFD)—with machine learning and deep learning classifiers. Baseline models demonstrated strong performance, with Random Forest achieving 97.41% accuracy ( \(\hbox {F1}=0.953\) ), XGBoost 98.20% ( \(\hbox {F1}=0.961\) ), and the Keras neural network 99.10% ( \(\hbox {F1}=0.980\) ). Transformer-based large language models (LLMs) further advanced this benchmark: Minos reached 100% accuracy (F1 = 0.993) and ViRanker 99.63% ( \(\hbox {F1}=0.982\) ). The ablation study confirmed that removing multi-sensor data (MODIS or VIIRS) or preprocessing components reduced performance by up to 3–4%, underscoring the synergy between satellite fusion and optimized pipelines. Ten-fold cross-validation validated the superior calibration and reliability of LLMs (Expected Calibration Error, ECE \(\approx\) 0.031), with efficient training ( \(\approx\) 78.5 s) and real-time inference ( \(<0.36\) s). Comparative analysis revealed that FlameGuard surpassed NASA’s FIRMS (75% accuracy, 3–6 h latency), MODIS fire alerts (70%, 4–6 h), and recent research (88–94%), achieving 99% accuracy, sub-hour detection latency, and IoT-ground-truth validation. All models were deployed within the FlameGuard smart map interface, enabling real-time visualization of fire risks, thermal anomalies, and confidence levels. The platform delivers interpretable, high-precision analytics to support proactive environmental monitoring and early response planning in fire-prone agricultural regions.
