Ilmu Komputer & AI editorial
Open AccessOA2025
FlameGuard: an AI-driven smart map for early detection and management of agricultural fires in Saudi Arabia
A geospatial AI framework integrating spectral rules, machine learning, and large language models for sub-hour fire detection
M. Aljaafari; Shaymaa E. Sorour· Scientific Reports· 2025· DOI 10.1038/s41598-025-28814-9
The core problem
Agricultural fires in arid and semi-arid regions, such as Saudi Arabia, pose significant environmental and economic threats. The increasing frequency of such fires necessitates advanced, scalable early detection systems. Traditional satellite-based fire monitoring systems, like NASA's FIRMS, offer moderate accuracy (75%) and latency of 3–6 hours, which is insufficient for timely response. This study introduces FlameGuard, an AI-driven smart map that leverages thermal satellite imagery from the FIRMS dataset and multi-model predictive architectures to classify fire confidence levels with high precision. The framework integrates a spectral rule-based model—Spectral Algorithm for Fire Detection (SAFD)—with machine learning and deep learning classifiers, aiming to provide interpretable, high-precision analytics for proactive fire management.
Innovation
The performance of FlameGuard models is summarized in Table 1. Baseline models achieved high accuracy: Random Forest 97.41% ( = 0.953), XGBoost 98.20% ( = 0.961), and Keras neural network 99.10% ( = 0.980). Transformer-based LLMs further improved performance: Minos reached 100% accuracy ( = 0.993) and ViRanker 99.63% ( = 0.982). An ablation study revealed that removing multi-sensor data (MODIS or VIIRS) or preprocessing components reduced performance by up to 3–4%, highlighting the importance of satellite fusion and optimized pipelines. Ten-fold cross-validation confirmed the superior calibration of LLMs with an Expected Calibration Error (ECE) of approximately 0.031. Training time was efficient (≈78.5 s) and real-time inference was achieved (<0.36 s). Comparative analysis showed 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 and sub-hour detection latency. The system also integrates IoT-ground-truth validation for enhanced reliability.
Agricultural fires in arid and semi-arid regions, such as Saudi Arabia, pose significant environmental and economic threats. The increasing frequency of such fires necessitates advanced, scalable early detection systems. Traditional satellite-based fire monitoring systems, like NASA's FIRMS, offer moderate accuracy (75%) and latency of 3–6 hours, which is insufficient for timely response. This study introduces FlameGuard, an AI-driven smart map that leverages thermal satellite imagery from the FIRMS dataset and multi-model predictive architectures to classify fire confidence levels with high precision. The framework integrates a spectral rule-based model—Spectral Algorithm for Fire Detection (SAFD)—with machine learning and deep learning classifiers, aiming to provide interpretable, high-precision analytics for proactive fire management.
FlameGuard utilizes thermal satellite imagery from MODIS and VIIRS sensors, obtained from the FIRMS dataset. The framework comprises three main components: (1) a spectral rule-based model (SAFD) for initial fire detection, (2) machine learning classifiers including Random Forest and XGBoost, and (3) deep learning models such as a Keras neural network and transformer-based large language models (Minos and ViRanker). The data preprocessing pipeline includes noise reduction, feature extraction, and fusion of multi-sensor data. The models were trained and evaluated using ten-fold cross-validation, with performance metrics including accuracy, -score, and Expected Calibration Error (ECE). The architecture is illustrated below:
Why it matters
The results demonstrate that FlameGuard significantly outperforms existing fire detection systems in terms of accuracy and latency. The integration of spectral rules with machine learning and deep learning models enables robust classification of fire confidence levels, even in challenging arid environments. The ablation study underscores the critical role of multi-sensor data fusion and preprocessing in achieving high performance. The use of transformer-based LLMs, such as Minos and ViRanker, provides state-of-the-art accuracy and well-calibrated confidence estimates, which are essential for decision-making. The FlameGuard smart map interface facilitates real-time visualization of fire risks, thermal anomalies, and confidence levels, empowering stakeholders to take proactive measures. Despite these advances, challenges remain, including the need for continuous model updates and integration with ground-based sensor networks. Future work will focus on expanding the framework to other regions and incorporating additional data sources. Overall, FlameGuard represents a significant step forward in AI-driven environmental monitoring and early response planning for agricultural fires.
Who should read this
CS practitioners and researchers
Opening member content…